| 1 | //===- GlobalOpt.cpp - Optimize Global Variables --------------------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This pass transforms simple global variables that never have their address |
| 10 | // taken. If obviously true, it marks read/write globals as constant, deletes |
| 11 | // variables only stored to, etc. |
| 12 | // |
| 13 | //===----------------------------------------------------------------------===// |
| 14 | |
| 15 | #include "llvm/Transforms/IPO/GlobalOpt.h" |
| 16 | #include "llvm/ADT/DenseMap.h" |
| 17 | #include "llvm/ADT/STLExtras.h" |
| 18 | #include "llvm/ADT/SmallPtrSet.h" |
| 19 | #include "llvm/ADT/SmallVector.h" |
| 20 | #include "llvm/ADT/Statistic.h" |
| 21 | #include "llvm/ADT/Twine.h" |
| 22 | #include "llvm/ADT/iterator_range.h" |
| 23 | #include "llvm/Analysis/BlockFrequencyInfo.h" |
| 24 | #include "llvm/Analysis/ConstantFolding.h" |
| 25 | #include "llvm/Analysis/MemoryBuiltins.h" |
| 26 | #include "llvm/Analysis/TargetLibraryInfo.h" |
| 27 | #include "llvm/Analysis/TargetTransformInfo.h" |
| 28 | #include "llvm/Analysis/ValueTracking.h" |
| 29 | #include "llvm/BinaryFormat/Dwarf.h" |
| 30 | #include "llvm/IR/Attributes.h" |
| 31 | #include "llvm/IR/BasicBlock.h" |
| 32 | #include "llvm/IR/CallingConv.h" |
| 33 | #include "llvm/IR/Constant.h" |
| 34 | #include "llvm/IR/Constants.h" |
| 35 | #include "llvm/IR/DataLayout.h" |
| 36 | #include "llvm/IR/DebugInfoMetadata.h" |
| 37 | #include "llvm/IR/DerivedTypes.h" |
| 38 | #include "llvm/IR/Dominators.h" |
| 39 | #include "llvm/IR/Function.h" |
| 40 | #include "llvm/IR/GlobalAlias.h" |
| 41 | #include "llvm/IR/GlobalValue.h" |
| 42 | #include "llvm/IR/GlobalVariable.h" |
| 43 | #include "llvm/IR/IRBuilder.h" |
| 44 | #include "llvm/IR/InstrTypes.h" |
| 45 | #include "llvm/IR/Instruction.h" |
| 46 | #include "llvm/IR/Instructions.h" |
| 47 | #include "llvm/IR/IntrinsicInst.h" |
| 48 | #include "llvm/IR/Module.h" |
| 49 | #include "llvm/IR/Operator.h" |
| 50 | #include "llvm/IR/ProfDataUtils.h" |
| 51 | #include "llvm/IR/Type.h" |
| 52 | #include "llvm/IR/Use.h" |
| 53 | #include "llvm/IR/User.h" |
| 54 | #include "llvm/IR/Value.h" |
| 55 | #include "llvm/IR/ValueHandle.h" |
| 56 | #include "llvm/Support/AtomicOrdering.h" |
| 57 | #include "llvm/Support/Casting.h" |
| 58 | #include "llvm/Support/CommandLine.h" |
| 59 | #include "llvm/Support/Debug.h" |
| 60 | #include "llvm/Support/ErrorHandling.h" |
| 61 | #include "llvm/Support/raw_ostream.h" |
| 62 | #include "llvm/Transforms/IPO.h" |
| 63 | #include "llvm/Transforms/Utils/CtorUtils.h" |
| 64 | #include "llvm/Transforms/Utils/Evaluator.h" |
| 65 | #include "llvm/Transforms/Utils/GlobalStatus.h" |
| 66 | #include "llvm/Transforms/Utils/Local.h" |
| 67 | #include <cassert> |
| 68 | #include <cstdint> |
| 69 | #include <optional> |
| 70 | #include <utility> |
| 71 | #include <vector> |
| 72 | |
| 73 | using namespace llvm; |
| 74 | |
| 75 | #define DEBUG_TYPE "globalopt" |
| 76 | |
| 77 | STATISTIC(NumMarked , "Number of globals marked constant" ); |
| 78 | STATISTIC(NumUnnamed , "Number of globals marked unnamed_addr" ); |
| 79 | STATISTIC(NumSRA , "Number of aggregate globals broken into scalars" ); |
| 80 | STATISTIC(NumSubstitute,"Number of globals with initializers stored into them" ); |
| 81 | STATISTIC(NumDeleted , "Number of globals deleted" ); |
| 82 | STATISTIC(NumGlobUses , "Number of global uses devirtualized" ); |
| 83 | STATISTIC(NumLocalized , "Number of globals localized" ); |
| 84 | STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans" ); |
| 85 | STATISTIC(NumFastCallFns , "Number of functions converted to fastcc" ); |
| 86 | STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated" ); |
| 87 | STATISTIC(NumNestRemoved , "Number of nest attributes removed" ); |
| 88 | STATISTIC(NumAliasesResolved, "Number of global aliases resolved" ); |
| 89 | STATISTIC(NumAliasesRemoved, "Number of global aliases eliminated" ); |
| 90 | STATISTIC(NumCXXDtorsRemoved, "Number of global C++ destructors removed" ); |
| 91 | STATISTIC(NumAtExitRemoved, "Number of atexit handlers removed" ); |
| 92 | STATISTIC(NumInternalFunc, "Number of internal functions" ); |
| 93 | STATISTIC(NumColdCC, "Number of functions marked coldcc" ); |
| 94 | STATISTIC(NumIFuncsResolved, "Number of statically resolved IFuncs" ); |
| 95 | STATISTIC(NumIFuncsDeleted, "Number of IFuncs removed" ); |
| 96 | |
| 97 | static cl::opt<bool> |
| 98 | OptimizeNonFMVCallers("optimize-non-fmv-callers" , |
| 99 | cl::desc("Statically resolve calls to versioned " |
| 100 | "functions from non-versioned callers." ), |
| 101 | cl::init(Val: true), cl::Hidden); |
| 102 | |
| 103 | static cl::opt<unsigned> MaxIFuncVersions( |
| 104 | "max-ifunc-versions" , cl::Hidden, cl::init(Val: 5), |
| 105 | cl::desc("Maximum number of caller/callee versions that is allowed for " |
| 106 | "using the expensive (cubic) static resolution algorithm." )); |
| 107 | |
| 108 | static cl::opt<bool> |
| 109 | EnableColdCCStressTest("enable-coldcc-stress-test" , |
| 110 | cl::desc("Enable stress test of coldcc by adding " |
| 111 | "calling conv to all internal functions." ), |
| 112 | cl::init(Val: false), cl::Hidden); |
| 113 | |
| 114 | static cl::opt<int> ColdCCRelFreq( |
| 115 | "coldcc-rel-freq" , cl::Hidden, cl::init(Val: 2), |
| 116 | cl::desc( |
| 117 | "Maximum block frequency, expressed as a percentage of caller's " |
| 118 | "entry frequency, for a call site to be considered cold for enabling " |
| 119 | "coldcc" )); |
| 120 | |
| 121 | /// Is this global variable possibly used by a leak checker as a root? If so, |
| 122 | /// we might not really want to eliminate the stores to it. |
| 123 | static bool isLeakCheckerRoot(GlobalVariable *GV) { |
| 124 | // A global variable is a root if it is a pointer, or could plausibly contain |
| 125 | // a pointer. There are two challenges; one is that we could have a struct |
| 126 | // the has an inner member which is a pointer. We recurse through the type to |
| 127 | // detect these (up to a point). The other is that we may actually be a union |
| 128 | // of a pointer and another type, and so our LLVM type is an integer which |
| 129 | // gets converted into a pointer, or our type is an [i8 x #] with a pointer |
| 130 | // potentially contained here. |
| 131 | |
| 132 | if (GV->hasPrivateLinkage()) |
| 133 | return false; |
| 134 | |
| 135 | SmallVector<Type *, 4> Types; |
| 136 | Types.push_back(Elt: GV->getValueType()); |
| 137 | |
| 138 | unsigned Limit = 20; |
| 139 | do { |
| 140 | Type *Ty = Types.pop_back_val(); |
| 141 | switch (Ty->getTypeID()) { |
| 142 | default: break; |
| 143 | case Type::PointerTyID: |
| 144 | return true; |
| 145 | case Type::FixedVectorTyID: |
| 146 | case Type::ScalableVectorTyID: |
| 147 | if (cast<VectorType>(Val: Ty)->getElementType()->isPointerTy()) |
| 148 | return true; |
| 149 | break; |
| 150 | case Type::ArrayTyID: |
| 151 | Types.push_back(Elt: cast<ArrayType>(Val: Ty)->getElementType()); |
| 152 | break; |
| 153 | case Type::StructTyID: { |
| 154 | StructType *STy = cast<StructType>(Val: Ty); |
| 155 | if (STy->isOpaque()) return true; |
| 156 | for (Type *InnerTy : STy->elements()) { |
| 157 | if (isa<PointerType>(Val: InnerTy)) return true; |
| 158 | if (isa<StructType>(Val: InnerTy) || isa<ArrayType>(Val: InnerTy) || |
| 159 | isa<VectorType>(Val: InnerTy)) |
| 160 | Types.push_back(Elt: InnerTy); |
| 161 | } |
| 162 | break; |
| 163 | } |
| 164 | } |
| 165 | if (--Limit == 0) return true; |
| 166 | } while (!Types.empty()); |
| 167 | return false; |
| 168 | } |
| 169 | |
| 170 | /// Given a value that is stored to a global but never read, determine whether |
| 171 | /// it's safe to remove the store and the chain of computation that feeds the |
| 172 | /// store. |
| 173 | static bool IsSafeComputationToRemove( |
| 174 | Value *V, function_ref<TargetLibraryInfo &(Function &)> GetTLI) { |
| 175 | do { |
| 176 | if (isa<Constant>(Val: V)) |
| 177 | return true; |
| 178 | if (!V->hasOneUse()) |
| 179 | return false; |
| 180 | if (isa<LoadInst>(Val: V) || isa<InvokeInst>(Val: V) || isa<Argument>(Val: V) || |
| 181 | isa<GlobalValue>(Val: V)) |
| 182 | return false; |
| 183 | if (isAllocationFn(V, GetTLI)) |
| 184 | return true; |
| 185 | |
| 186 | Instruction *I = cast<Instruction>(Val: V); |
| 187 | if (I->mayHaveSideEffects()) |
| 188 | return false; |
| 189 | if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Val: I)) { |
| 190 | if (!GEP->hasAllConstantIndices()) |
| 191 | return false; |
| 192 | } else if (I->getNumOperands() != 1) { |
| 193 | return false; |
| 194 | } |
| 195 | |
| 196 | V = I->getOperand(i: 0); |
| 197 | } while (true); |
| 198 | } |
| 199 | |
| 200 | /// This GV is a pointer root. Loop over all users of the global and clean up |
| 201 | /// any that obviously don't assign the global a value that isn't dynamically |
| 202 | /// allocated. |
| 203 | static bool |
| 204 | CleanupPointerRootUsers(GlobalVariable *GV, |
| 205 | function_ref<TargetLibraryInfo &(Function &)> GetTLI) { |
| 206 | // A brief explanation of leak checkers. The goal is to find bugs where |
| 207 | // pointers are forgotten, causing an accumulating growth in memory |
| 208 | // usage over time. The common strategy for leak checkers is to explicitly |
| 209 | // allow the memory pointed to by globals at exit. This is popular because it |
| 210 | // also solves another problem where the main thread of a C++ program may shut |
| 211 | // down before other threads that are still expecting to use those globals. To |
| 212 | // handle that case, we expect the program may create a singleton and never |
| 213 | // destroy it. |
| 214 | |
| 215 | bool Changed = false; |
| 216 | |
| 217 | // If Dead[n].first is the only use of a malloc result, we can delete its |
| 218 | // chain of computation and the store to the global in Dead[n].second. |
| 219 | SmallVector<std::pair<Instruction *, Instruction *>, 32> Dead; |
| 220 | |
| 221 | SmallVector<User *> Worklist(GV->users()); |
| 222 | // Constants can't be pointers to dynamically allocated memory. |
| 223 | while (!Worklist.empty()) { |
| 224 | User *U = Worklist.pop_back_val(); |
| 225 | if (StoreInst *SI = dyn_cast<StoreInst>(Val: U)) { |
| 226 | Value *V = SI->getValueOperand(); |
| 227 | if (isa<Constant>(Val: V)) { |
| 228 | Changed = true; |
| 229 | SI->eraseFromParent(); |
| 230 | } else if (Instruction *I = dyn_cast<Instruction>(Val: V)) { |
| 231 | if (I->hasOneUse()) |
| 232 | Dead.push_back(Elt: std::make_pair(x&: I, y&: SI)); |
| 233 | } |
| 234 | } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(Val: U)) { |
| 235 | if (isa<Constant>(Val: MSI->getValue())) { |
| 236 | Changed = true; |
| 237 | MSI->eraseFromParent(); |
| 238 | } else if (Instruction *I = dyn_cast<Instruction>(Val: MSI->getValue())) { |
| 239 | if (I->hasOneUse()) |
| 240 | Dead.push_back(Elt: std::make_pair(x&: I, y&: MSI)); |
| 241 | } |
| 242 | } else if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(Val: U)) { |
| 243 | GlobalVariable *MemSrc = dyn_cast<GlobalVariable>(Val: MTI->getSource()); |
| 244 | if (MemSrc && MemSrc->isConstant()) { |
| 245 | Changed = true; |
| 246 | MTI->eraseFromParent(); |
| 247 | } else if (Instruction *I = dyn_cast<Instruction>(Val: MTI->getSource())) { |
| 248 | if (I->hasOneUse()) |
| 249 | Dead.push_back(Elt: std::make_pair(x&: I, y&: MTI)); |
| 250 | } |
| 251 | } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Val: U)) { |
| 252 | if (isa<GEPOperator>(Val: CE)) |
| 253 | append_range(C&: Worklist, R: CE->users()); |
| 254 | } |
| 255 | } |
| 256 | |
| 257 | for (const auto &[Inst, Store] : Dead) { |
| 258 | if (IsSafeComputationToRemove(V: Inst, GetTLI)) { |
| 259 | Store->eraseFromParent(); |
| 260 | Instruction *I = Inst; |
| 261 | do { |
| 262 | if (isAllocationFn(V: I, GetTLI)) |
| 263 | break; |
| 264 | Instruction *J = dyn_cast<Instruction>(Val: I->getOperand(i: 0)); |
| 265 | if (!J) |
| 266 | break; |
| 267 | I->eraseFromParent(); |
| 268 | I = J; |
| 269 | } while (true); |
| 270 | I->eraseFromParent(); |
| 271 | Changed = true; |
| 272 | } |
| 273 | } |
| 274 | |
| 275 | GV->removeDeadConstantUsers(); |
| 276 | return Changed; |
| 277 | } |
| 278 | |
| 279 | /// We just marked GV constant. Loop over all users of the global, cleaning up |
| 280 | /// the obvious ones. This is largely just a quick scan over the use list to |
| 281 | /// clean up the easy and obvious cruft. This returns true if it made a change. |
| 282 | static bool CleanupConstantGlobalUsers(GlobalVariable *GV, |
| 283 | const DataLayout &DL) { |
| 284 | Constant *Init = GV->getInitializer(); |
| 285 | SmallVector<User *, 8> WorkList(GV->users()); |
| 286 | SmallPtrSet<User *, 8> Visited; |
| 287 | bool Changed = false; |
| 288 | |
| 289 | SmallVector<WeakTrackingVH> MaybeDeadInsts; |
| 290 | auto EraseFromParent = [&](Instruction *I) { |
| 291 | for (Value *Op : I->operands()) |
| 292 | if (auto *OpI = dyn_cast<Instruction>(Val: Op)) |
| 293 | MaybeDeadInsts.push_back(Elt: OpI); |
| 294 | I->eraseFromParent(); |
| 295 | Changed = true; |
| 296 | }; |
| 297 | while (!WorkList.empty()) { |
| 298 | User *U = WorkList.pop_back_val(); |
| 299 | if (!Visited.insert(Ptr: U).second) |
| 300 | continue; |
| 301 | |
| 302 | if (auto *BO = dyn_cast<BitCastOperator>(Val: U)) |
| 303 | append_range(C&: WorkList, R: BO->users()); |
| 304 | if (auto *ASC = dyn_cast<AddrSpaceCastOperator>(Val: U)) |
| 305 | append_range(C&: WorkList, R: ASC->users()); |
| 306 | else if (auto *GEP = dyn_cast<GEPOperator>(Val: U)) |
| 307 | append_range(C&: WorkList, R: GEP->users()); |
| 308 | else if (auto *LI = dyn_cast<LoadInst>(Val: U)) { |
| 309 | // A load from a uniform value is always the same, regardless of any |
| 310 | // applied offset. |
| 311 | Type *Ty = LI->getType(); |
| 312 | if (Constant *Res = ConstantFoldLoadFromUniformValue(C: Init, Ty, DL)) { |
| 313 | LI->replaceAllUsesWith(V: Res); |
| 314 | EraseFromParent(LI); |
| 315 | continue; |
| 316 | } |
| 317 | |
| 318 | Value *PtrOp = LI->getPointerOperand(); |
| 319 | APInt Offset(DL.getIndexTypeSizeInBits(Ty: PtrOp->getType()), 0); |
| 320 | PtrOp = PtrOp->stripAndAccumulateConstantOffsets( |
| 321 | DL, Offset, /* AllowNonInbounds */ true); |
| 322 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: PtrOp)) { |
| 323 | if (II->getIntrinsicID() == Intrinsic::threadlocal_address) |
| 324 | PtrOp = II->getArgOperand(i: 0); |
| 325 | } |
| 326 | if (PtrOp == GV) { |
| 327 | if (auto *Value = ConstantFoldLoadFromConst(C: Init, Ty, Offset, DL)) { |
| 328 | LI->replaceAllUsesWith(V: Value); |
| 329 | EraseFromParent(LI); |
| 330 | } |
| 331 | } |
| 332 | } else if (StoreInst *SI = dyn_cast<StoreInst>(Val: U)) { |
| 333 | // Store must be unreachable or storing Init into the global. |
| 334 | EraseFromParent(SI); |
| 335 | } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(Val: U)) { // memset/cpy/mv |
| 336 | if (getUnderlyingObject(V: MI->getRawDest()) == GV) |
| 337 | EraseFromParent(MI); |
| 338 | } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: U)) { |
| 339 | if (II->getIntrinsicID() == Intrinsic::threadlocal_address) |
| 340 | append_range(C&: WorkList, R: II->users()); |
| 341 | } |
| 342 | } |
| 343 | |
| 344 | Changed |= |
| 345 | RecursivelyDeleteTriviallyDeadInstructionsPermissive(DeadInsts&: MaybeDeadInsts); |
| 346 | GV->removeDeadConstantUsers(); |
| 347 | return Changed; |
| 348 | } |
| 349 | |
| 350 | /// Part of the global at a specific offset, which is only accessed through |
| 351 | /// loads and stores with the given type. |
| 352 | struct GlobalPart { |
| 353 | Type *Ty; |
| 354 | Constant *Initializer = nullptr; |
| 355 | bool IsLoaded = false; |
| 356 | bool IsStored = false; |
| 357 | }; |
| 358 | |
| 359 | /// Look at all uses of the global and determine which (offset, type) pairs it |
| 360 | /// can be split into. |
| 361 | static bool collectSRATypes(DenseMap<uint64_t, GlobalPart> &Parts, |
| 362 | GlobalVariable *GV, const DataLayout &DL) { |
| 363 | SmallVector<Use *, 16> Worklist; |
| 364 | SmallPtrSet<Use *, 16> Visited; |
| 365 | auto AppendUses = [&](Value *V) { |
| 366 | for (Use &U : V->uses()) |
| 367 | if (Visited.insert(Ptr: &U).second) |
| 368 | Worklist.push_back(Elt: &U); |
| 369 | }; |
| 370 | AppendUses(GV); |
| 371 | while (!Worklist.empty()) { |
| 372 | Use *U = Worklist.pop_back_val(); |
| 373 | User *V = U->getUser(); |
| 374 | |
| 375 | auto *GEP = dyn_cast<GEPOperator>(Val: V); |
| 376 | if (isa<BitCastOperator>(Val: V) || isa<AddrSpaceCastOperator>(Val: V) || |
| 377 | (GEP && GEP->hasAllConstantIndices())) { |
| 378 | AppendUses(V); |
| 379 | continue; |
| 380 | } |
| 381 | |
| 382 | if (Value *Ptr = getLoadStorePointerOperand(V)) { |
| 383 | // This is storing the global address into somewhere, not storing into |
| 384 | // the global. |
| 385 | if (isa<StoreInst>(Val: V) && U->getOperandNo() == 0) |
| 386 | return false; |
| 387 | |
| 388 | APInt Offset(DL.getIndexTypeSizeInBits(Ty: Ptr->getType()), 0); |
| 389 | Ptr = Ptr->stripAndAccumulateConstantOffsets(DL, Offset, |
| 390 | /* AllowNonInbounds */ true); |
| 391 | if (Ptr != GV || Offset.getActiveBits() >= 64) |
| 392 | return false; |
| 393 | |
| 394 | // TODO: We currently require that all accesses at a given offset must |
| 395 | // use the same type. This could be relaxed. |
| 396 | Type *Ty = getLoadStoreType(I: V); |
| 397 | const auto &[It, Inserted] = |
| 398 | Parts.try_emplace(Key: Offset.getZExtValue(), Args: GlobalPart{.Ty: Ty}); |
| 399 | if (Ty != It->second.Ty) |
| 400 | return false; |
| 401 | |
| 402 | if (Inserted) { |
| 403 | It->second.Initializer = |
| 404 | ConstantFoldLoadFromConst(C: GV->getInitializer(), Ty, Offset, DL); |
| 405 | if (!It->second.Initializer) { |
| 406 | LLVM_DEBUG(dbgs() << "Global SRA: Failed to evaluate initializer of " |
| 407 | << *GV << " with type " << *Ty << " at offset " |
| 408 | << Offset.getZExtValue()); |
| 409 | return false; |
| 410 | } |
| 411 | } |
| 412 | |
| 413 | // Scalable types not currently supported. |
| 414 | if (Ty->isScalableTy()) |
| 415 | return false; |
| 416 | |
| 417 | auto IsStored = [](Value *V, Constant *Initializer) { |
| 418 | auto *SI = dyn_cast<StoreInst>(Val: V); |
| 419 | if (!SI) |
| 420 | return false; |
| 421 | |
| 422 | Constant *StoredConst = dyn_cast<Constant>(Val: SI->getOperand(i_nocapture: 0)); |
| 423 | if (!StoredConst) |
| 424 | return true; |
| 425 | |
| 426 | // Don't consider stores that only write the initializer value. |
| 427 | return Initializer != StoredConst; |
| 428 | }; |
| 429 | |
| 430 | It->second.IsLoaded |= isa<LoadInst>(Val: V); |
| 431 | It->second.IsStored |= IsStored(V, It->second.Initializer); |
| 432 | continue; |
| 433 | } |
| 434 | |
| 435 | // Ignore dead constant users. |
| 436 | if (auto *C = dyn_cast<Constant>(Val: V)) { |
| 437 | if (!isSafeToDestroyConstant(C)) |
| 438 | return false; |
| 439 | continue; |
| 440 | } |
| 441 | |
| 442 | // Unknown user. |
| 443 | return false; |
| 444 | } |
| 445 | |
| 446 | return true; |
| 447 | } |
| 448 | |
| 449 | /// Copy over the debug info for a variable to its SRA replacements. |
| 450 | static void transferSRADebugInfo(GlobalVariable *GV, GlobalVariable *NGV, |
| 451 | uint64_t FragmentOffsetInBits, |
| 452 | uint64_t FragmentSizeInBits, |
| 453 | uint64_t VarSize) { |
| 454 | SmallVector<DIGlobalVariableExpression *, 1> GVs; |
| 455 | GV->getDebugInfo(GVs); |
| 456 | for (auto *GVE : GVs) { |
| 457 | DIVariable *Var = GVE->getVariable(); |
| 458 | DIExpression *Expr = GVE->getExpression(); |
| 459 | int64_t CurVarOffsetInBytes = 0; |
| 460 | uint64_t CurVarOffsetInBits = 0; |
| 461 | uint64_t FragmentEndInBits = FragmentOffsetInBits + FragmentSizeInBits; |
| 462 | |
| 463 | // Calculate the offset (Bytes), Continue if unknown. |
| 464 | if (!Expr->extractIfOffset(Offset&: CurVarOffsetInBytes)) |
| 465 | continue; |
| 466 | |
| 467 | // Ignore negative offset. |
| 468 | if (CurVarOffsetInBytes < 0) |
| 469 | continue; |
| 470 | |
| 471 | // Convert offset to bits. |
| 472 | CurVarOffsetInBits = CHAR_BIT * (uint64_t)CurVarOffsetInBytes; |
| 473 | |
| 474 | // Current var starts after the fragment, ignore. |
| 475 | if (CurVarOffsetInBits >= FragmentEndInBits) |
| 476 | continue; |
| 477 | |
| 478 | uint64_t CurVarSize = Var->getType()->getSizeInBits(); |
| 479 | uint64_t CurVarEndInBits = CurVarOffsetInBits + CurVarSize; |
| 480 | // Current variable ends before start of fragment, ignore. |
| 481 | if (CurVarSize != 0 && /* CurVarSize is known */ |
| 482 | CurVarEndInBits <= FragmentOffsetInBits) |
| 483 | continue; |
| 484 | |
| 485 | // Current variable fits in (not greater than) the fragment, |
| 486 | // does not need fragment expression. |
| 487 | if (CurVarSize != 0 && /* CurVarSize is known */ |
| 488 | CurVarOffsetInBits >= FragmentOffsetInBits && |
| 489 | CurVarEndInBits <= FragmentEndInBits) { |
| 490 | uint64_t CurVarOffsetInFragment = |
| 491 | (CurVarOffsetInBits - FragmentOffsetInBits) / 8; |
| 492 | if (CurVarOffsetInFragment != 0) |
| 493 | Expr = DIExpression::get(Context&: Expr->getContext(), Elements: {dwarf::DW_OP_plus_uconst, |
| 494 | CurVarOffsetInFragment}); |
| 495 | else |
| 496 | Expr = DIExpression::get(Context&: Expr->getContext(), Elements: {}); |
| 497 | auto *NGVE = |
| 498 | DIGlobalVariableExpression::get(Context&: GVE->getContext(), Variable: Var, Expression: Expr); |
| 499 | NGV->addDebugInfo(GV: NGVE); |
| 500 | continue; |
| 501 | } |
| 502 | // Current variable does not fit in single fragment, |
| 503 | // emit a fragment expression. |
| 504 | if (FragmentSizeInBits < VarSize) { |
| 505 | if (CurVarOffsetInBits > FragmentOffsetInBits) |
| 506 | continue; |
| 507 | uint64_t CurVarFragmentOffsetInBits = |
| 508 | FragmentOffsetInBits - CurVarOffsetInBits; |
| 509 | uint64_t CurVarFragmentSizeInBits = FragmentSizeInBits; |
| 510 | if (CurVarSize != 0 && CurVarEndInBits < FragmentEndInBits) |
| 511 | CurVarFragmentSizeInBits -= (FragmentEndInBits - CurVarEndInBits); |
| 512 | if (CurVarOffsetInBits) |
| 513 | Expr = DIExpression::get(Context&: Expr->getContext(), Elements: {}); |
| 514 | if (auto E = DIExpression::createFragmentExpression( |
| 515 | Expr, OffsetInBits: CurVarFragmentOffsetInBits, SizeInBits: CurVarFragmentSizeInBits)) |
| 516 | Expr = *E; |
| 517 | else |
| 518 | continue; |
| 519 | } |
| 520 | auto *NGVE = DIGlobalVariableExpression::get(Context&: GVE->getContext(), Variable: Var, Expression: Expr); |
| 521 | NGV->addDebugInfo(GV: NGVE); |
| 522 | } |
| 523 | } |
| 524 | |
| 525 | /// Perform scalar replacement of aggregates on the specified global variable. |
| 526 | /// This opens the door for other optimizations by exposing the behavior of the |
| 527 | /// program in a more fine-grained way. We have determined that this |
| 528 | /// transformation is safe already. We return the first global variable we |
| 529 | /// insert so that the caller can reprocess it. |
| 530 | static GlobalVariable *SRAGlobal(GlobalVariable *GV, const DataLayout &DL) { |
| 531 | assert(GV->hasLocalLinkage()); |
| 532 | |
| 533 | // Collect types to split into. |
| 534 | DenseMap<uint64_t, GlobalPart> Parts; |
| 535 | if (!collectSRATypes(Parts, GV, DL) || Parts.empty()) |
| 536 | return nullptr; |
| 537 | |
| 538 | // Make sure we don't SRA back to the same type. |
| 539 | if (Parts.size() == 1 && Parts.begin()->second.Ty == GV->getValueType()) |
| 540 | return nullptr; |
| 541 | |
| 542 | // Don't perform SRA if we would have to split into many globals. Ignore |
| 543 | // parts that are either only loaded or only stored, because we expect them |
| 544 | // to be optimized away. |
| 545 | unsigned NumParts = count_if(Range&: Parts, P: [](const auto &Pair) { |
| 546 | return Pair.second.IsLoaded && Pair.second.IsStored; |
| 547 | }); |
| 548 | if (NumParts > 16) |
| 549 | return nullptr; |
| 550 | |
| 551 | // Sort by offset. |
| 552 | SmallVector<std::tuple<uint64_t, Type *, Constant *>, 16> TypesVector; |
| 553 | for (const auto &Pair : Parts) { |
| 554 | TypesVector.push_back( |
| 555 | Elt: {Pair.first, Pair.second.Ty, Pair.second.Initializer}); |
| 556 | } |
| 557 | sort(C&: TypesVector, Comp: llvm::less_first()); |
| 558 | |
| 559 | // Check that the types are non-overlapping. |
| 560 | uint64_t Offset = 0; |
| 561 | for (const auto &[OffsetForTy, Ty, _] : TypesVector) { |
| 562 | // Overlaps with previous type. |
| 563 | if (OffsetForTy < Offset) |
| 564 | return nullptr; |
| 565 | |
| 566 | Offset = OffsetForTy + DL.getTypeAllocSize(Ty); |
| 567 | } |
| 568 | |
| 569 | // Some accesses go beyond the end of the global, don't bother. |
| 570 | if (Offset > GV->getGlobalSize(DL)) |
| 571 | return nullptr; |
| 572 | |
| 573 | LLVM_DEBUG(dbgs() << "PERFORMING GLOBAL SRA ON: " << *GV << "\n" ); |
| 574 | |
| 575 | // Get the alignment of the global, either explicit or target-specific. |
| 576 | Align StartAlignment = |
| 577 | DL.getValueOrABITypeAlignment(Alignment: GV->getAlign(), Ty: GV->getValueType()); |
| 578 | uint64_t VarSize = DL.getTypeSizeInBits(Ty: GV->getValueType()); |
| 579 | |
| 580 | // Create replacement globals. |
| 581 | DenseMap<uint64_t, GlobalVariable *> NewGlobals; |
| 582 | unsigned NameSuffix = 0; |
| 583 | for (auto &[OffsetForTy, Ty, Initializer] : TypesVector) { |
| 584 | GlobalVariable *NGV = new GlobalVariable( |
| 585 | *GV->getParent(), Ty, false, GlobalVariable::InternalLinkage, |
| 586 | Initializer, GV->getName() + "." + Twine(NameSuffix++), GV, |
| 587 | GV->getThreadLocalMode(), GV->getAddressSpace()); |
| 588 | // Start out by copying attributes from the original, including alignment. |
| 589 | NGV->copyAttributesFrom(Src: GV); |
| 590 | NewGlobals.insert(KV: {OffsetForTy, NGV}); |
| 591 | |
| 592 | // Calculate the known alignment of the field. If the original aggregate |
| 593 | // had 256 byte alignment for example, then the element at a given offset |
| 594 | // may also have a known alignment, and something might depend on that: |
| 595 | // propagate info to each field. |
| 596 | Align NewAlign = commonAlignment(A: StartAlignment, Offset: OffsetForTy); |
| 597 | NGV->setAlignment(NewAlign); |
| 598 | |
| 599 | // Copy over the debug info for the variable. |
| 600 | transferSRADebugInfo(GV, NGV, FragmentOffsetInBits: OffsetForTy * 8, |
| 601 | FragmentSizeInBits: DL.getTypeAllocSizeInBits(Ty), VarSize); |
| 602 | } |
| 603 | |
| 604 | // Replace uses of the original global with uses of the new global. |
| 605 | SmallVector<Value *, 16> Worklist; |
| 606 | SmallPtrSet<Value *, 16> Visited; |
| 607 | SmallVector<WeakTrackingVH, 16> DeadInsts; |
| 608 | auto AppendUsers = [&](Value *V) { |
| 609 | for (User *U : V->users()) |
| 610 | if (Visited.insert(Ptr: U).second) |
| 611 | Worklist.push_back(Elt: U); |
| 612 | }; |
| 613 | AppendUsers(GV); |
| 614 | while (!Worklist.empty()) { |
| 615 | Value *V = Worklist.pop_back_val(); |
| 616 | if (isa<BitCastOperator>(Val: V) || isa<AddrSpaceCastOperator>(Val: V) || |
| 617 | isa<GEPOperator>(Val: V)) { |
| 618 | AppendUsers(V); |
| 619 | if (isa<Instruction>(Val: V)) |
| 620 | DeadInsts.push_back(Elt: V); |
| 621 | continue; |
| 622 | } |
| 623 | |
| 624 | if (Value *Ptr = getLoadStorePointerOperand(V)) { |
| 625 | APInt Offset(DL.getIndexTypeSizeInBits(Ty: Ptr->getType()), 0); |
| 626 | Ptr = Ptr->stripAndAccumulateConstantOffsets(DL, Offset, |
| 627 | /* AllowNonInbounds */ true); |
| 628 | assert(Ptr == GV && "Load/store must be from/to global" ); |
| 629 | GlobalVariable *NGV = NewGlobals[Offset.getZExtValue()]; |
| 630 | assert(NGV && "Must have replacement global for this offset" ); |
| 631 | |
| 632 | // Update the pointer operand and recalculate alignment. |
| 633 | Align PrefAlign = DL.getPrefTypeAlign(Ty: getLoadStoreType(I: V)); |
| 634 | Align NewAlign = |
| 635 | getOrEnforceKnownAlignment(V: NGV, PrefAlign, DL, CxtI: cast<Instruction>(Val: V)); |
| 636 | |
| 637 | if (auto *LI = dyn_cast<LoadInst>(Val: V)) { |
| 638 | LI->setOperand(i_nocapture: 0, Val_nocapture: NGV); |
| 639 | LI->setAlignment(NewAlign); |
| 640 | } else { |
| 641 | auto *SI = cast<StoreInst>(Val: V); |
| 642 | SI->setOperand(i_nocapture: 1, Val_nocapture: NGV); |
| 643 | SI->setAlignment(NewAlign); |
| 644 | } |
| 645 | continue; |
| 646 | } |
| 647 | |
| 648 | assert(isa<Constant>(V) && isSafeToDestroyConstant(cast<Constant>(V)) && |
| 649 | "Other users can only be dead constants" ); |
| 650 | } |
| 651 | |
| 652 | // Delete old instructions and global. |
| 653 | RecursivelyDeleteTriviallyDeadInstructions(DeadInsts); |
| 654 | GV->removeDeadConstantUsers(); |
| 655 | GV->eraseFromParent(); |
| 656 | ++NumSRA; |
| 657 | |
| 658 | assert(NewGlobals.size() > 0); |
| 659 | return NewGlobals.begin()->second; |
| 660 | } |
| 661 | |
| 662 | /// Return true if all users of the specified value will trap if the value is |
| 663 | /// dynamically null. PHIs keeps track of any phi nodes we've seen to avoid |
| 664 | /// reprocessing them. |
| 665 | static bool AllUsesOfValueWillTrapIfNull(const Value *V, |
| 666 | SmallPtrSetImpl<const PHINode*> &PHIs) { |
| 667 | for (const User *U : V->users()) { |
| 668 | if (const Instruction *I = dyn_cast<Instruction>(Val: U)) { |
| 669 | // If null pointer is considered valid, then all uses are non-trapping. |
| 670 | // Non address-space 0 globals have already been pruned by the caller. |
| 671 | if (NullPointerIsDefined(F: I->getFunction())) |
| 672 | return false; |
| 673 | } |
| 674 | if (isa<LoadInst>(Val: U)) { |
| 675 | // Will trap. |
| 676 | } else if (const StoreInst *SI = dyn_cast<StoreInst>(Val: U)) { |
| 677 | if (SI->getOperand(i_nocapture: 0) == V) { |
| 678 | return false; // Storing the value. |
| 679 | } |
| 680 | } else if (const CallInst *CI = dyn_cast<CallInst>(Val: U)) { |
| 681 | if (CI->getCalledOperand() != V) { |
| 682 | return false; // Not calling the ptr |
| 683 | } |
| 684 | } else if (const InvokeInst *II = dyn_cast<InvokeInst>(Val: U)) { |
| 685 | if (II->getCalledOperand() != V) { |
| 686 | return false; // Not calling the ptr |
| 687 | } |
| 688 | } else if (const AddrSpaceCastInst *CI = dyn_cast<AddrSpaceCastInst>(Val: U)) { |
| 689 | if (!AllUsesOfValueWillTrapIfNull(V: CI, PHIs)) |
| 690 | return false; |
| 691 | } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Val: U)) { |
| 692 | if (!AllUsesOfValueWillTrapIfNull(V: GEPI, PHIs)) return false; |
| 693 | } else if (const PHINode *PN = dyn_cast<PHINode>(Val: U)) { |
| 694 | // If we've already seen this phi node, ignore it, it has already been |
| 695 | // checked. |
| 696 | if (PHIs.insert(Ptr: PN).second && !AllUsesOfValueWillTrapIfNull(V: PN, PHIs)) |
| 697 | return false; |
| 698 | } else if (isa<ICmpInst>(Val: U) && |
| 699 | !ICmpInst::isSigned(predicate: cast<ICmpInst>(Val: U)->getPredicate()) && |
| 700 | isa<LoadInst>(Val: U->getOperand(i: 0)) && |
| 701 | isa<ConstantPointerNull>(Val: U->getOperand(i: 1))) { |
| 702 | assert(isa<GlobalValue>(cast<LoadInst>(U->getOperand(0)) |
| 703 | ->getPointerOperand() |
| 704 | ->stripPointerCasts()) && |
| 705 | "Should be GlobalVariable" ); |
| 706 | // This and only this kind of non-signed ICmpInst is to be replaced with |
| 707 | // the comparing of the value of the created global init bool later in |
| 708 | // optimizeGlobalAddressOfAllocation for the global variable. |
| 709 | } else { |
| 710 | return false; |
| 711 | } |
| 712 | } |
| 713 | return true; |
| 714 | } |
| 715 | |
| 716 | /// Return true if all uses of any loads from GV will trap if the loaded value |
| 717 | /// is null. Note that this also permits comparisons of the loaded value |
| 718 | /// against null, as a special case. |
| 719 | static bool allUsesOfLoadedValueWillTrapIfNull(const GlobalVariable *GV) { |
| 720 | SmallVector<const Value *, 4> Worklist; |
| 721 | Worklist.push_back(Elt: GV); |
| 722 | while (!Worklist.empty()) { |
| 723 | const Value *P = Worklist.pop_back_val(); |
| 724 | for (const auto *U : P->users()) { |
| 725 | if (auto *LI = dyn_cast<LoadInst>(Val: U)) { |
| 726 | if (!LI->isSimple()) |
| 727 | return false; |
| 728 | SmallPtrSet<const PHINode *, 8> PHIs; |
| 729 | if (!AllUsesOfValueWillTrapIfNull(V: LI, PHIs)) |
| 730 | return false; |
| 731 | } else if (auto *SI = dyn_cast<StoreInst>(Val: U)) { |
| 732 | if (!SI->isSimple()) |
| 733 | return false; |
| 734 | // Ignore stores to the global. |
| 735 | if (SI->getPointerOperand() != P) |
| 736 | return false; |
| 737 | } else if (auto *CE = dyn_cast<ConstantExpr>(Val: U)) { |
| 738 | if (CE->stripPointerCasts() != GV) |
| 739 | return false; |
| 740 | // Check further the ConstantExpr. |
| 741 | Worklist.push_back(Elt: CE); |
| 742 | } else { |
| 743 | // We don't know or understand this user, bail out. |
| 744 | return false; |
| 745 | } |
| 746 | } |
| 747 | } |
| 748 | |
| 749 | return true; |
| 750 | } |
| 751 | |
| 752 | /// Get all the loads/store uses for global variable \p GV. |
| 753 | static void allUsesOfLoadAndStores(GlobalVariable *GV, |
| 754 | SmallVector<Value *, 4> &Uses) { |
| 755 | SmallVector<Value *, 4> Worklist; |
| 756 | Worklist.push_back(Elt: GV); |
| 757 | while (!Worklist.empty()) { |
| 758 | auto *P = Worklist.pop_back_val(); |
| 759 | for (auto *U : P->users()) { |
| 760 | if (auto *CE = dyn_cast<ConstantExpr>(Val: U)) { |
| 761 | Worklist.push_back(Elt: CE); |
| 762 | continue; |
| 763 | } |
| 764 | |
| 765 | assert((isa<LoadInst>(U) || isa<StoreInst>(U)) && |
| 766 | "Expect only load or store instructions" ); |
| 767 | Uses.push_back(Elt: U); |
| 768 | } |
| 769 | } |
| 770 | } |
| 771 | |
| 772 | static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) { |
| 773 | bool Changed = false; |
| 774 | for (auto UI = V->user_begin(), E = V->user_end(); UI != E; ) { |
| 775 | Instruction *I = cast<Instruction>(Val: *UI++); |
| 776 | // Uses are non-trapping if null pointer is considered valid. |
| 777 | // Non address-space 0 globals are already pruned by the caller. |
| 778 | if (NullPointerIsDefined(F: I->getFunction())) |
| 779 | return false; |
| 780 | if (LoadInst *LI = dyn_cast<LoadInst>(Val: I)) { |
| 781 | LI->setOperand(i_nocapture: 0, Val_nocapture: NewV); |
| 782 | Changed = true; |
| 783 | } else if (StoreInst *SI = dyn_cast<StoreInst>(Val: I)) { |
| 784 | if (SI->getOperand(i_nocapture: 1) == V) { |
| 785 | SI->setOperand(i_nocapture: 1, Val_nocapture: NewV); |
| 786 | Changed = true; |
| 787 | } |
| 788 | } else if (isa<CallInst>(Val: I) || isa<InvokeInst>(Val: I)) { |
| 789 | CallBase *CB = cast<CallBase>(Val: I); |
| 790 | if (CB->getCalledOperand() == V) { |
| 791 | // Calling through the pointer! Turn into a direct call, but be careful |
| 792 | // that the pointer is not also being passed as an argument. |
| 793 | CB->setCalledOperand(NewV); |
| 794 | Changed = true; |
| 795 | bool PassedAsArg = false; |
| 796 | for (unsigned i = 0, e = CB->arg_size(); i != e; ++i) |
| 797 | if (CB->getArgOperand(i) == V) { |
| 798 | PassedAsArg = true; |
| 799 | CB->setArgOperand(i, v: NewV); |
| 800 | } |
| 801 | |
| 802 | if (PassedAsArg) { |
| 803 | // Being passed as an argument also. Be careful to not invalidate UI! |
| 804 | UI = V->user_begin(); |
| 805 | } |
| 806 | } |
| 807 | } else if (AddrSpaceCastInst *CI = dyn_cast<AddrSpaceCastInst>(Val: I)) { |
| 808 | Changed |= OptimizeAwayTrappingUsesOfValue( |
| 809 | V: CI, NewV: ConstantExpr::getAddrSpaceCast(C: NewV, Ty: CI->getType())); |
| 810 | if (CI->use_empty()) { |
| 811 | Changed = true; |
| 812 | CI->eraseFromParent(); |
| 813 | } |
| 814 | } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Val: I)) { |
| 815 | // Should handle GEP here. |
| 816 | SmallVector<Constant*, 8> Idxs; |
| 817 | Idxs.reserve(N: GEPI->getNumOperands()-1); |
| 818 | for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end(); |
| 819 | i != e; ++i) |
| 820 | if (Constant *C = dyn_cast<Constant>(Val&: *i)) |
| 821 | Idxs.push_back(Elt: C); |
| 822 | else |
| 823 | break; |
| 824 | if (Idxs.size() == GEPI->getNumOperands()-1) |
| 825 | Changed |= OptimizeAwayTrappingUsesOfValue( |
| 826 | V: GEPI, NewV: ConstantExpr::getGetElementPtr(Ty: GEPI->getSourceElementType(), |
| 827 | C: NewV, IdxList: Idxs)); |
| 828 | if (GEPI->use_empty()) { |
| 829 | Changed = true; |
| 830 | GEPI->eraseFromParent(); |
| 831 | } |
| 832 | } |
| 833 | } |
| 834 | |
| 835 | return Changed; |
| 836 | } |
| 837 | |
| 838 | /// The specified global has only one non-null value stored into it. If there |
| 839 | /// are uses of the loaded value that would trap if the loaded value is |
| 840 | /// dynamically null, then we know that they cannot be reachable with a null |
| 841 | /// optimize away the load. |
| 842 | static bool OptimizeAwayTrappingUsesOfLoads( |
| 843 | GlobalVariable *GV, Constant *LV, const DataLayout &DL, |
| 844 | function_ref<TargetLibraryInfo &(Function &)> GetTLI) { |
| 845 | bool Changed = false; |
| 846 | |
| 847 | // Keep track of whether we are able to remove all the uses of the global |
| 848 | // other than the store that defines it. |
| 849 | bool AllNonStoreUsesGone = true; |
| 850 | |
| 851 | // Replace all uses of loads with uses of uses of the stored value. |
| 852 | for (User *GlobalUser : llvm::make_early_inc_range(Range: GV->users())) { |
| 853 | if (LoadInst *LI = dyn_cast<LoadInst>(Val: GlobalUser)) { |
| 854 | Changed |= OptimizeAwayTrappingUsesOfValue(V: LI, NewV: LV); |
| 855 | // If we were able to delete all uses of the loads |
| 856 | if (LI->use_empty()) { |
| 857 | LI->eraseFromParent(); |
| 858 | Changed = true; |
| 859 | } else { |
| 860 | AllNonStoreUsesGone = false; |
| 861 | } |
| 862 | } else if (isa<StoreInst>(Val: GlobalUser)) { |
| 863 | // Ignore the store that stores "LV" to the global. |
| 864 | assert(GlobalUser->getOperand(1) == GV && |
| 865 | "Must be storing *to* the global" ); |
| 866 | } else { |
| 867 | AllNonStoreUsesGone = false; |
| 868 | |
| 869 | // If we get here we could have other crazy uses that are transitively |
| 870 | // loaded. |
| 871 | assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) || |
| 872 | isa<ConstantExpr>(GlobalUser) || isa<CmpInst>(GlobalUser) || |
| 873 | isa<BitCastInst>(GlobalUser) || |
| 874 | isa<GetElementPtrInst>(GlobalUser) || |
| 875 | isa<AddrSpaceCastInst>(GlobalUser)) && |
| 876 | "Only expect load and stores!" ); |
| 877 | } |
| 878 | } |
| 879 | |
| 880 | if (Changed) { |
| 881 | LLVM_DEBUG(dbgs() << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GV |
| 882 | << "\n" ); |
| 883 | ++NumGlobUses; |
| 884 | } |
| 885 | |
| 886 | // If we nuked all of the loads, then none of the stores are needed either, |
| 887 | // nor is the global. |
| 888 | if (AllNonStoreUsesGone) { |
| 889 | if (isLeakCheckerRoot(GV)) { |
| 890 | Changed |= CleanupPointerRootUsers(GV, GetTLI); |
| 891 | } else { |
| 892 | Changed = true; |
| 893 | CleanupConstantGlobalUsers(GV, DL); |
| 894 | } |
| 895 | if (GV->use_empty()) { |
| 896 | LLVM_DEBUG(dbgs() << " *** GLOBAL NOW DEAD!\n" ); |
| 897 | Changed = true; |
| 898 | GV->eraseFromParent(); |
| 899 | ++NumDeleted; |
| 900 | } |
| 901 | } |
| 902 | return Changed; |
| 903 | } |
| 904 | |
| 905 | /// Walk the use list of V, constant folding all of the instructions that are |
| 906 | /// foldable. |
| 907 | static void ConstantPropUsersOf(Value *V, const DataLayout &DL, |
| 908 | TargetLibraryInfo *TLI) { |
| 909 | for (Value::user_iterator UI = V->user_begin(), E = V->user_end(); UI != E; ) |
| 910 | if (Instruction *I = dyn_cast<Instruction>(Val: *UI++)) |
| 911 | if (Constant *NewC = ConstantFoldInstruction(I, DL, TLI)) { |
| 912 | I->replaceAllUsesWith(V: NewC); |
| 913 | |
| 914 | // Advance UI to the next non-I use to avoid invalidating it! |
| 915 | // Instructions could multiply use V. |
| 916 | while (UI != E && *UI == I) |
| 917 | ++UI; |
| 918 | if (isInstructionTriviallyDead(I, TLI)) |
| 919 | I->eraseFromParent(); |
| 920 | } |
| 921 | } |
| 922 | |
| 923 | /// This function takes the specified global variable, and transforms the |
| 924 | /// program as if it always contained the result of the specified malloc. |
| 925 | /// Because it is always the result of the specified malloc, there is no reason |
| 926 | /// to actually DO the malloc. Instead, turn the malloc into a global, and any |
| 927 | /// loads of GV as uses of the new global. |
| 928 | static GlobalVariable * |
| 929 | OptimizeGlobalAddressOfAllocation(GlobalVariable *GV, CallInst *CI, |
| 930 | uint64_t AllocSize, Constant *InitVal, |
| 931 | const DataLayout &DL, |
| 932 | TargetLibraryInfo *TLI) { |
| 933 | LLVM_DEBUG(errs() << "PROMOTING GLOBAL: " << *GV << " CALL = " << *CI |
| 934 | << '\n'); |
| 935 | |
| 936 | // Create global of type [AllocSize x i8]. |
| 937 | Type *GlobalType = ArrayType::get(ElementType: Type::getInt8Ty(C&: GV->getContext()), |
| 938 | NumElements: AllocSize); |
| 939 | |
| 940 | // Create the new global variable. The contents of the allocated memory is |
| 941 | // undefined initially, so initialize with an undef value. |
| 942 | GlobalVariable *NewGV = new GlobalVariable( |
| 943 | *GV->getParent(), GlobalType, false, GlobalValue::InternalLinkage, |
| 944 | UndefValue::get(T: GlobalType), GV->getName() + ".body" , nullptr, |
| 945 | GV->getThreadLocalMode()); |
| 946 | |
| 947 | // Initialize the global at the point of the original call. Note that this |
| 948 | // is a different point from the initialization referred to below for the |
| 949 | // nullability handling. Sublety: We have not proven the original global was |
| 950 | // only initialized once. As such, we can not fold this into the initializer |
| 951 | // of the new global as may need to re-init the storage multiple times. |
| 952 | if (!isa<UndefValue>(Val: InitVal)) { |
| 953 | IRBuilder<> Builder(CI->getNextNode()); |
| 954 | // TODO: Use alignment above if align!=1 |
| 955 | Builder.CreateMemSet(Ptr: NewGV, Val: InitVal, Size: AllocSize, Align: std::nullopt); |
| 956 | } |
| 957 | |
| 958 | // Update users of the allocation to use the new global instead. |
| 959 | CI->replaceAllUsesWith(V: NewGV); |
| 960 | |
| 961 | // If there is a comparison against null, we will insert a global bool to |
| 962 | // keep track of whether the global was initialized yet or not. |
| 963 | GlobalVariable *InitBool = new GlobalVariable( |
| 964 | Type::getInt1Ty(C&: GV->getContext()), false, GlobalValue::InternalLinkage, |
| 965 | ConstantInt::getFalse(Context&: GV->getContext()), GV->getName() + ".init" , |
| 966 | GV->getThreadLocalMode(), GV->getAddressSpace()); |
| 967 | bool InitBoolUsed = false; |
| 968 | |
| 969 | // Loop over all instruction uses of GV, processing them in turn. |
| 970 | SmallVector<Value *, 4> Guses; |
| 971 | allUsesOfLoadAndStores(GV, Uses&: Guses); |
| 972 | for (auto *U : Guses) { |
| 973 | if (StoreInst *SI = dyn_cast<StoreInst>(Val: U)) { |
| 974 | // The global is initialized when the store to it occurs. If the stored |
| 975 | // value is null value, the global bool is set to false, otherwise true. |
| 976 | auto *NewSI = new StoreInst( |
| 977 | ConstantInt::getBool(Context&: GV->getContext(), V: !isa<ConstantPointerNull>( |
| 978 | Val: SI->getValueOperand())), |
| 979 | InitBool, false, Align(1), SI->getOrdering(), SI->getSyncScopeID(), |
| 980 | SI->getIterator()); |
| 981 | NewSI->setDebugLoc(SI->getDebugLoc()); |
| 982 | SI->eraseFromParent(); |
| 983 | continue; |
| 984 | } |
| 985 | |
| 986 | LoadInst *LI = cast<LoadInst>(Val: U); |
| 987 | while (!LI->use_empty()) { |
| 988 | Use &LoadUse = *LI->use_begin(); |
| 989 | ICmpInst *ICI = dyn_cast<ICmpInst>(Val: LoadUse.getUser()); |
| 990 | if (!ICI) { |
| 991 | LoadUse.set(NewGV); |
| 992 | continue; |
| 993 | } |
| 994 | |
| 995 | // Replace the cmp X, 0 with a use of the bool value. |
| 996 | Value *LV = new LoadInst(InitBool->getValueType(), InitBool, |
| 997 | InitBool->getName() + ".val" , false, Align(1), |
| 998 | LI->getOrdering(), LI->getSyncScopeID(), |
| 999 | LI->getIterator()); |
| 1000 | // FIXME: Should we use the DebugLoc of the load used by the predicate, or |
| 1001 | // the predicate? The load seems most appropriate, but there's an argument |
| 1002 | // that the new load does not represent the old load, but is simply a |
| 1003 | // component of recomputing the predicate. |
| 1004 | cast<LoadInst>(Val: LV)->setDebugLoc(LI->getDebugLoc()); |
| 1005 | InitBoolUsed = true; |
| 1006 | switch (ICI->getPredicate()) { |
| 1007 | default: llvm_unreachable("Unknown ICmp Predicate!" ); |
| 1008 | case ICmpInst::ICMP_ULT: // X < null -> always false |
| 1009 | LV = ConstantInt::getFalse(Context&: GV->getContext()); |
| 1010 | break; |
| 1011 | case ICmpInst::ICMP_UGE: // X >= null -> always true |
| 1012 | LV = ConstantInt::getTrue(Context&: GV->getContext()); |
| 1013 | break; |
| 1014 | case ICmpInst::ICMP_ULE: |
| 1015 | case ICmpInst::ICMP_EQ: |
| 1016 | LV = BinaryOperator::CreateNot(Op: LV, Name: "notinit" , InsertBefore: ICI->getIterator()); |
| 1017 | cast<BinaryOperator>(Val: LV)->setDebugLoc(ICI->getDebugLoc()); |
| 1018 | break; |
| 1019 | case ICmpInst::ICMP_NE: |
| 1020 | case ICmpInst::ICMP_UGT: |
| 1021 | break; // no change. |
| 1022 | } |
| 1023 | ICI->replaceAllUsesWith(V: LV); |
| 1024 | ICI->eraseFromParent(); |
| 1025 | } |
| 1026 | LI->eraseFromParent(); |
| 1027 | } |
| 1028 | |
| 1029 | // If the initialization boolean was used, insert it, otherwise delete it. |
| 1030 | if (!InitBoolUsed) { |
| 1031 | while (!InitBool->use_empty()) // Delete initializations |
| 1032 | cast<StoreInst>(Val: InitBool->user_back())->eraseFromParent(); |
| 1033 | delete InitBool; |
| 1034 | } else |
| 1035 | GV->getParent()->insertGlobalVariable(Where: GV->getIterator(), GV: InitBool); |
| 1036 | |
| 1037 | // Now the GV is dead, nuke it and the allocation.. |
| 1038 | GV->eraseFromParent(); |
| 1039 | CI->eraseFromParent(); |
| 1040 | |
| 1041 | // To further other optimizations, loop over all users of NewGV and try to |
| 1042 | // constant prop them. This will promote GEP instructions with constant |
| 1043 | // indices into GEP constant-exprs, which will allow global-opt to hack on it. |
| 1044 | ConstantPropUsersOf(V: NewGV, DL, TLI); |
| 1045 | |
| 1046 | return NewGV; |
| 1047 | } |
| 1048 | |
| 1049 | /// Scan the use-list of GV checking to make sure that there are no complex uses |
| 1050 | /// of GV. We permit simple things like dereferencing the pointer, but not |
| 1051 | /// storing through the address, unless it is to the specified global. |
| 1052 | static bool |
| 1053 | valueIsOnlyUsedLocallyOrStoredToOneGlobal(const CallInst *CI, |
| 1054 | const GlobalVariable *GV) { |
| 1055 | SmallPtrSet<const Value *, 4> Visited; |
| 1056 | SmallVector<const Value *, 4> Worklist; |
| 1057 | Worklist.push_back(Elt: CI); |
| 1058 | |
| 1059 | while (!Worklist.empty()) { |
| 1060 | const Value *V = Worklist.pop_back_val(); |
| 1061 | if (!Visited.insert(Ptr: V).second) |
| 1062 | continue; |
| 1063 | |
| 1064 | for (const Use &VUse : V->uses()) { |
| 1065 | const User *U = VUse.getUser(); |
| 1066 | if (isa<LoadInst>(Val: U) || isa<CmpInst>(Val: U)) |
| 1067 | continue; // Fine, ignore. |
| 1068 | |
| 1069 | if (auto *SI = dyn_cast<StoreInst>(Val: U)) { |
| 1070 | if (SI->getValueOperand() == V && |
| 1071 | SI->getPointerOperand()->stripPointerCasts() != GV) |
| 1072 | return false; // Storing the pointer not into GV... bad. |
| 1073 | continue; // Otherwise, storing through it, or storing into GV... fine. |
| 1074 | } |
| 1075 | |
| 1076 | if (auto *GEPI = dyn_cast<GetElementPtrInst>(Val: U)) { |
| 1077 | Worklist.push_back(Elt: GEPI); |
| 1078 | continue; |
| 1079 | } |
| 1080 | |
| 1081 | return false; |
| 1082 | } |
| 1083 | } |
| 1084 | |
| 1085 | return true; |
| 1086 | } |
| 1087 | |
| 1088 | /// If we have a global that is only initialized with a fixed size allocation |
| 1089 | /// try to transform the program to use global memory instead of heap |
| 1090 | /// allocated memory. This eliminates dynamic allocation, avoids an indirection |
| 1091 | /// accessing the data, and exposes the resultant global to further GlobalOpt. |
| 1092 | static bool tryToOptimizeStoreOfAllocationToGlobal(GlobalVariable *GV, |
| 1093 | CallInst *CI, |
| 1094 | const DataLayout &DL, |
| 1095 | TargetLibraryInfo *TLI) { |
| 1096 | if (!isRemovableAlloc(V: CI, TLI)) |
| 1097 | // Must be able to remove the call when we get done.. |
| 1098 | return false; |
| 1099 | |
| 1100 | Type *Int8Ty = Type::getInt8Ty(C&: CI->getFunction()->getContext()); |
| 1101 | Constant *InitVal = getInitialValueOfAllocation(V: CI, TLI, Ty: Int8Ty); |
| 1102 | if (!InitVal) |
| 1103 | // Must be able to emit a memset for initialization |
| 1104 | return false; |
| 1105 | |
| 1106 | uint64_t AllocSize; |
| 1107 | if (!getObjectSize(Ptr: CI, Size&: AllocSize, DL, TLI, Opts: ObjectSizeOpts())) |
| 1108 | return false; |
| 1109 | |
| 1110 | // Restrict this transformation to only working on small allocations |
| 1111 | // (2048 bytes currently), as we don't want to introduce a 16M global or |
| 1112 | // something. |
| 1113 | if (AllocSize >= 2048) |
| 1114 | return false; |
| 1115 | |
| 1116 | // We can't optimize this global unless all uses of it are *known* to be |
| 1117 | // of the malloc value, not of the null initializer value (consider a use |
| 1118 | // that compares the global's value against zero to see if the malloc has |
| 1119 | // been reached). To do this, we check to see if all uses of the global |
| 1120 | // would trap if the global were null: this proves that they must all |
| 1121 | // happen after the malloc. |
| 1122 | if (!allUsesOfLoadedValueWillTrapIfNull(GV)) |
| 1123 | return false; |
| 1124 | |
| 1125 | // We can't optimize this if the malloc itself is used in a complex way, |
| 1126 | // for example, being stored into multiple globals. This allows the |
| 1127 | // malloc to be stored into the specified global, loaded, gep, icmp'd. |
| 1128 | // These are all things we could transform to using the global for. |
| 1129 | if (!valueIsOnlyUsedLocallyOrStoredToOneGlobal(CI, GV)) |
| 1130 | return false; |
| 1131 | |
| 1132 | OptimizeGlobalAddressOfAllocation(GV, CI, AllocSize, InitVal, DL, TLI); |
| 1133 | return true; |
| 1134 | } |
| 1135 | |
| 1136 | // Try to optimize globals based on the knowledge that only one value (besides |
| 1137 | // its initializer) is ever stored to the global. |
| 1138 | static bool |
| 1139 | optimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal, |
| 1140 | const DataLayout &DL, |
| 1141 | function_ref<TargetLibraryInfo &(Function &)> GetTLI) { |
| 1142 | // If we are dealing with a pointer global that is initialized to null and |
| 1143 | // only has one (non-null) value stored into it, then we can optimize any |
| 1144 | // users of the loaded value (often calls and loads) that would trap if the |
| 1145 | // value was null. |
| 1146 | if (GV->getInitializer()->getType()->isPointerTy() && |
| 1147 | GV->getInitializer()->isNullValue() && |
| 1148 | StoredOnceVal->getType()->isPointerTy() && |
| 1149 | !NullPointerIsDefined( |
| 1150 | F: nullptr /* F */, |
| 1151 | AS: GV->getInitializer()->getType()->getPointerAddressSpace())) { |
| 1152 | if (Constant *SOVC = dyn_cast<Constant>(Val: StoredOnceVal)) { |
| 1153 | // Optimize away any trapping uses of the loaded value. |
| 1154 | if (OptimizeAwayTrappingUsesOfLoads(GV, LV: SOVC, DL, GetTLI)) |
| 1155 | return true; |
| 1156 | } else if (isAllocationFn(V: StoredOnceVal, GetTLI)) { |
| 1157 | if (auto *CI = dyn_cast<CallInst>(Val: StoredOnceVal)) { |
| 1158 | auto *TLI = &GetTLI(*CI->getFunction()); |
| 1159 | if (tryToOptimizeStoreOfAllocationToGlobal(GV, CI, DL, TLI)) |
| 1160 | return true; |
| 1161 | } |
| 1162 | } |
| 1163 | } |
| 1164 | |
| 1165 | return false; |
| 1166 | } |
| 1167 | |
| 1168 | /// At this point, we have learned that the only two values ever stored into GV |
| 1169 | /// are its initializer and OtherVal. See if we can shrink the global into a |
| 1170 | /// boolean and select between the two values whenever it is used. This exposes |
| 1171 | /// the values to other scalar optimizations. |
| 1172 | static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) { |
| 1173 | Type *GVElType = GV->getValueType(); |
| 1174 | |
| 1175 | // If GVElType is already i1, it is already shrunk. If the type of the GV is |
| 1176 | // an FP value, pointer or vector, don't do this optimization because a select |
| 1177 | // between them is very expensive and unlikely to lead to later |
| 1178 | // simplification. In these cases, we typically end up with "cond ? v1 : v2" |
| 1179 | // where v1 and v2 both require constant pool loads, a big loss. |
| 1180 | if (GVElType == Type::getInt1Ty(C&: GV->getContext()) || |
| 1181 | GVElType->isFloatingPointTy() || |
| 1182 | GVElType->isPointerTy() || GVElType->isVectorTy()) |
| 1183 | return false; |
| 1184 | |
| 1185 | // Walk the use list of the global seeing if all the uses are load or store. |
| 1186 | // If there is anything else, bail out. |
| 1187 | for (User *U : GV->users()) { |
| 1188 | if (!isa<LoadInst>(Val: U) && !isa<StoreInst>(Val: U)) |
| 1189 | return false; |
| 1190 | if (getLoadStoreType(I: U) != GVElType) |
| 1191 | return false; |
| 1192 | } |
| 1193 | |
| 1194 | LLVM_DEBUG(dbgs() << " *** SHRINKING TO BOOL: " << *GV << "\n" ); |
| 1195 | |
| 1196 | // Create the new global, initializing it to false. |
| 1197 | GlobalVariable *NewGV = new GlobalVariable(Type::getInt1Ty(C&: GV->getContext()), |
| 1198 | false, |
| 1199 | GlobalValue::InternalLinkage, |
| 1200 | ConstantInt::getFalse(Context&: GV->getContext()), |
| 1201 | GV->getName()+".b" , |
| 1202 | GV->getThreadLocalMode(), |
| 1203 | GV->getType()->getAddressSpace()); |
| 1204 | NewGV->copyAttributesFrom(Src: GV); |
| 1205 | GV->getParent()->insertGlobalVariable(Where: GV->getIterator(), GV: NewGV); |
| 1206 | |
| 1207 | Constant *InitVal = GV->getInitializer(); |
| 1208 | assert(InitVal->getType() != Type::getInt1Ty(GV->getContext()) && |
| 1209 | "No reason to shrink to bool!" ); |
| 1210 | |
| 1211 | SmallVector<DIGlobalVariableExpression *, 1> GVs; |
| 1212 | GV->getDebugInfo(GVs); |
| 1213 | |
| 1214 | // If initialized to zero and storing one into the global, we can use a cast |
| 1215 | // instead of a select to synthesize the desired value. |
| 1216 | bool IsOneZero = false; |
| 1217 | bool EmitOneOrZero = true; |
| 1218 | auto *CI = dyn_cast<ConstantInt>(Val: OtherVal); |
| 1219 | if (CI && CI->getValue().getActiveBits() <= 64) { |
| 1220 | IsOneZero = InitVal->isNullValue() && CI->isOne(); |
| 1221 | |
| 1222 | auto *CIInit = dyn_cast<ConstantInt>(Val: GV->getInitializer()); |
| 1223 | if (CIInit && CIInit->getValue().getActiveBits() <= 64) { |
| 1224 | uint64_t ValInit = CIInit->getZExtValue(); |
| 1225 | uint64_t ValOther = CI->getZExtValue(); |
| 1226 | uint64_t ValMinus = ValOther - ValInit; |
| 1227 | |
| 1228 | for(auto *GVe : GVs){ |
| 1229 | DIGlobalVariable *DGV = GVe->getVariable(); |
| 1230 | DIExpression *E = GVe->getExpression(); |
| 1231 | const DataLayout &DL = GV->getDataLayout(); |
| 1232 | unsigned SizeInOctets = NewGV->getGlobalSize(DL); |
| 1233 | |
| 1234 | // It is expected that the address of global optimized variable is on |
| 1235 | // top of the stack. After optimization, value of that variable will |
| 1236 | // be ether 0 for initial value or 1 for other value. The following |
| 1237 | // expression should return constant integer value depending on the |
| 1238 | // value at global object address: |
| 1239 | // val * (ValOther - ValInit) + ValInit: |
| 1240 | // DW_OP_deref DW_OP_constu <ValMinus> |
| 1241 | // DW_OP_mul DW_OP_constu <ValInit> DW_OP_plus DW_OP_stack_value |
| 1242 | SmallVector<uint64_t, 12> Ops = { |
| 1243 | dwarf::DW_OP_deref_size, SizeInOctets, |
| 1244 | dwarf::DW_OP_constu, ValMinus, |
| 1245 | dwarf::DW_OP_mul, dwarf::DW_OP_constu, ValInit, |
| 1246 | dwarf::DW_OP_plus}; |
| 1247 | bool WithStackValue = true; |
| 1248 | E = DIExpression::prependOpcodes(Expr: E, Ops, StackValue: WithStackValue); |
| 1249 | DIGlobalVariableExpression *DGVE = |
| 1250 | DIGlobalVariableExpression::get(Context&: NewGV->getContext(), Variable: DGV, Expression: E); |
| 1251 | NewGV->addDebugInfo(GV: DGVE); |
| 1252 | } |
| 1253 | EmitOneOrZero = false; |
| 1254 | } |
| 1255 | } |
| 1256 | |
| 1257 | if (EmitOneOrZero) { |
| 1258 | // FIXME: This will only emit address for debugger on which will |
| 1259 | // be written only 0 or 1. |
| 1260 | for(auto *GV : GVs) |
| 1261 | NewGV->addDebugInfo(GV); |
| 1262 | } |
| 1263 | |
| 1264 | while (!GV->use_empty()) { |
| 1265 | Instruction *UI = cast<Instruction>(Val: GV->user_back()); |
| 1266 | if (StoreInst *SI = dyn_cast<StoreInst>(Val: UI)) { |
| 1267 | // Change the store into a boolean store. |
| 1268 | bool StoringOther = SI->getOperand(i_nocapture: 0) == OtherVal; |
| 1269 | // Only do this if we weren't storing a loaded value. |
| 1270 | Value *StoreVal; |
| 1271 | if (StoringOther || SI->getOperand(i_nocapture: 0) == InitVal) { |
| 1272 | StoreVal = ConstantInt::get(Ty: Type::getInt1Ty(C&: GV->getContext()), |
| 1273 | V: StoringOther); |
| 1274 | } else { |
| 1275 | // Otherwise, we are storing a previously loaded copy. To do this, |
| 1276 | // change the copy from copying the original value to just copying the |
| 1277 | // bool. |
| 1278 | Instruction *StoredVal = cast<Instruction>(Val: SI->getOperand(i_nocapture: 0)); |
| 1279 | |
| 1280 | // If we've already replaced the input, StoredVal will be a cast or |
| 1281 | // select instruction. If not, it will be a load of the original |
| 1282 | // global. |
| 1283 | if (LoadInst *LI = dyn_cast<LoadInst>(Val: StoredVal)) { |
| 1284 | assert(LI->getOperand(0) == GV && "Not a copy!" ); |
| 1285 | // Insert a new load, to preserve the saved value. |
| 1286 | StoreVal = |
| 1287 | new LoadInst(NewGV->getValueType(), NewGV, LI->getName() + ".b" , |
| 1288 | false, Align(1), LI->getOrdering(), |
| 1289 | LI->getSyncScopeID(), LI->getIterator()); |
| 1290 | cast<LoadInst>(Val: StoreVal)->setDebugLoc(LI->getDebugLoc()); |
| 1291 | } else { |
| 1292 | assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) && |
| 1293 | "This is not a form that we understand!" ); |
| 1294 | StoreVal = StoredVal->getOperand(i: 0); |
| 1295 | assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!" ); |
| 1296 | } |
| 1297 | } |
| 1298 | StoreInst *NSI = |
| 1299 | new StoreInst(StoreVal, NewGV, false, Align(1), SI->getOrdering(), |
| 1300 | SI->getSyncScopeID(), SI->getIterator()); |
| 1301 | NSI->setDebugLoc(SI->getDebugLoc()); |
| 1302 | } else { |
| 1303 | // Change the load into a load of bool then a select. |
| 1304 | LoadInst *LI = cast<LoadInst>(Val: UI); |
| 1305 | LoadInst *NLI = new LoadInst( |
| 1306 | NewGV->getValueType(), NewGV, LI->getName() + ".b" , false, Align(1), |
| 1307 | LI->getOrdering(), LI->getSyncScopeID(), LI->getIterator()); |
| 1308 | Instruction *NSI; |
| 1309 | if (IsOneZero) |
| 1310 | NSI = new ZExtInst(NLI, LI->getType(), "" , LI->getIterator()); |
| 1311 | else { |
| 1312 | NSI = SelectInst::Create(C: NLI, S1: OtherVal, S2: InitVal, NameStr: "" , InsertBefore: LI->getIterator()); |
| 1313 | setExplicitlyUnknownBranchWeightsIfProfiled(I&: *NSI, DEBUG_TYPE); |
| 1314 | } |
| 1315 | NSI->takeName(V: LI); |
| 1316 | // Since LI is split into two instructions, NLI and NSI both inherit the |
| 1317 | // same DebugLoc |
| 1318 | NLI->setDebugLoc(LI->getDebugLoc()); |
| 1319 | NSI->setDebugLoc(LI->getDebugLoc()); |
| 1320 | LI->replaceAllUsesWith(V: NSI); |
| 1321 | } |
| 1322 | UI->eraseFromParent(); |
| 1323 | } |
| 1324 | |
| 1325 | // Retain the name of the old global variable. People who are debugging their |
| 1326 | // programs may expect these variables to be named the same. |
| 1327 | NewGV->takeName(V: GV); |
| 1328 | GV->eraseFromParent(); |
| 1329 | return true; |
| 1330 | } |
| 1331 | |
| 1332 | static bool |
| 1333 | deleteIfDead(GlobalValue &GV, |
| 1334 | SmallPtrSetImpl<const Comdat *> &NotDiscardableComdats, |
| 1335 | function_ref<void(Function &)> DeleteFnCallback = nullptr) { |
| 1336 | GV.removeDeadConstantUsers(); |
| 1337 | |
| 1338 | if (!GV.isDiscardableIfUnused() && !GV.isDeclaration()) |
| 1339 | return false; |
| 1340 | |
| 1341 | if (const Comdat *C = GV.getComdat()) |
| 1342 | if (!GV.hasLocalLinkage() && NotDiscardableComdats.count(Ptr: C)) |
| 1343 | return false; |
| 1344 | |
| 1345 | bool Dead; |
| 1346 | if (auto *F = dyn_cast<Function>(Val: &GV)) |
| 1347 | Dead = (F->isDeclaration() && F->use_empty()) || F->isDefTriviallyDead(); |
| 1348 | else |
| 1349 | Dead = GV.use_empty(); |
| 1350 | if (!Dead) |
| 1351 | return false; |
| 1352 | |
| 1353 | LLVM_DEBUG(dbgs() << "GLOBAL DEAD: " << GV << "\n" ); |
| 1354 | if (auto *F = dyn_cast<Function>(Val: &GV)) { |
| 1355 | if (DeleteFnCallback) |
| 1356 | DeleteFnCallback(*F); |
| 1357 | } |
| 1358 | ReplaceableMetadataImpl::SalvageDebugInfo(C: GV); |
| 1359 | GV.eraseFromParent(); |
| 1360 | ++NumDeleted; |
| 1361 | return true; |
| 1362 | } |
| 1363 | |
| 1364 | static bool isPointerValueDeadOnEntryToFunction( |
| 1365 | const Function *F, GlobalValue *GV, |
| 1366 | function_ref<DominatorTree &(Function &)> LookupDomTree) { |
| 1367 | // Find all uses of GV. We expect them all to be in F, and if we can't |
| 1368 | // identify any of the uses we bail out. |
| 1369 | // |
| 1370 | // On each of these uses, identify if the memory that GV points to is |
| 1371 | // used/required/live at the start of the function. If it is not, for example |
| 1372 | // if the first thing the function does is store to the GV, the GV can |
| 1373 | // possibly be demoted. |
| 1374 | // |
| 1375 | // We don't do an exhaustive search for memory operations - simply look |
| 1376 | // through bitcasts as they're quite common and benign. |
| 1377 | const DataLayout &DL = GV->getDataLayout(); |
| 1378 | SmallVector<LoadInst *, 4> Loads; |
| 1379 | SmallVector<StoreInst *, 4> Stores; |
| 1380 | for (auto *U : GV->users()) { |
| 1381 | Instruction *I = dyn_cast<Instruction>(Val: U); |
| 1382 | if (!I) |
| 1383 | return false; |
| 1384 | assert(I->getParent()->getParent() == F); |
| 1385 | |
| 1386 | if (auto *LI = dyn_cast<LoadInst>(Val: I)) |
| 1387 | Loads.push_back(Elt: LI); |
| 1388 | else if (auto *SI = dyn_cast<StoreInst>(Val: I)) |
| 1389 | Stores.push_back(Elt: SI); |
| 1390 | else |
| 1391 | return false; |
| 1392 | } |
| 1393 | |
| 1394 | // We have identified all uses of GV into loads and stores. Now check if all |
| 1395 | // of them are known not to depend on the value of the global at the function |
| 1396 | // entry point. We do this by ensuring that every load is dominated by at |
| 1397 | // least one store. |
| 1398 | auto &DT = LookupDomTree(*const_cast<Function *>(F)); |
| 1399 | |
| 1400 | // The below check is quadratic. Check we're not going to do too many tests. |
| 1401 | // FIXME: Even though this will always have worst-case quadratic time, we |
| 1402 | // could put effort into minimizing the average time by putting stores that |
| 1403 | // have been shown to dominate at least one load at the beginning of the |
| 1404 | // Stores array, making subsequent dominance checks more likely to succeed |
| 1405 | // early. |
| 1406 | // |
| 1407 | // The threshold here is fairly large because global->local demotion is a |
| 1408 | // very powerful optimization should it fire. |
| 1409 | const unsigned Threshold = 100; |
| 1410 | if (Loads.size() * Stores.size() > Threshold) |
| 1411 | return false; |
| 1412 | |
| 1413 | for (auto *L : Loads) { |
| 1414 | auto *LTy = L->getType(); |
| 1415 | if (none_of(Range&: Stores, P: [&](const StoreInst *S) { |
| 1416 | auto *STy = S->getValueOperand()->getType(); |
| 1417 | // The load is only dominated by the store if DomTree says so |
| 1418 | // and the number of bits loaded in L is less than or equal to |
| 1419 | // the number of bits stored in S. |
| 1420 | return DT.dominates(Def: S, User: L) && |
| 1421 | DL.getTypeStoreSize(Ty: LTy).getFixedValue() <= |
| 1422 | DL.getTypeStoreSize(Ty: STy).getFixedValue(); |
| 1423 | })) |
| 1424 | return false; |
| 1425 | } |
| 1426 | // All loads have known dependences inside F, so the global can be localized. |
| 1427 | return true; |
| 1428 | } |
| 1429 | |
| 1430 | // For a global variable with one store, if the store dominates any loads, |
| 1431 | // those loads will always load the stored value (as opposed to the |
| 1432 | // initializer), even in the presence of recursion. |
| 1433 | static bool forwardStoredOnceStore( |
| 1434 | GlobalVariable *GV, const StoreInst *StoredOnceStore, |
| 1435 | function_ref<DominatorTree &(Function &)> LookupDomTree) { |
| 1436 | const Value *StoredOnceValue = StoredOnceStore->getValueOperand(); |
| 1437 | // We can do this optimization for non-constants in nosync + norecurse |
| 1438 | // functions, but globals used in exactly one norecurse functions are already |
| 1439 | // promoted to an alloca. |
| 1440 | if (!isa<Constant>(Val: StoredOnceValue)) |
| 1441 | return false; |
| 1442 | const Function *F = StoredOnceStore->getFunction(); |
| 1443 | SmallVector<LoadInst *> Loads; |
| 1444 | for (User *U : GV->users()) { |
| 1445 | if (auto *LI = dyn_cast<LoadInst>(Val: U)) { |
| 1446 | if (LI->getFunction() == F && |
| 1447 | LI->getType() == StoredOnceValue->getType() && LI->isSimple()) |
| 1448 | Loads.push_back(Elt: LI); |
| 1449 | } |
| 1450 | } |
| 1451 | // Only compute DT if we have any loads to examine. |
| 1452 | bool MadeChange = false; |
| 1453 | if (!Loads.empty()) { |
| 1454 | auto &DT = LookupDomTree(*const_cast<Function *>(F)); |
| 1455 | for (auto *LI : Loads) { |
| 1456 | if (DT.dominates(Def: StoredOnceStore, User: LI)) { |
| 1457 | LI->replaceAllUsesWith(V: const_cast<Value *>(StoredOnceValue)); |
| 1458 | LI->eraseFromParent(); |
| 1459 | MadeChange = true; |
| 1460 | } |
| 1461 | } |
| 1462 | } |
| 1463 | return MadeChange; |
| 1464 | } |
| 1465 | |
| 1466 | /// Analyze the specified global variable and optimize |
| 1467 | /// it if possible. If we make a change, return true. |
| 1468 | static bool |
| 1469 | processInternalGlobal(GlobalVariable *GV, const GlobalStatus &GS, |
| 1470 | function_ref<TargetTransformInfo &(Function &)> GetTTI, |
| 1471 | function_ref<TargetLibraryInfo &(Function &)> GetTLI, |
| 1472 | function_ref<DominatorTree &(Function &)> LookupDomTree) { |
| 1473 | auto &DL = GV->getDataLayout(); |
| 1474 | // If this is a first class global and has only one accessing function and |
| 1475 | // this function is non-recursive, we replace the global with a local alloca |
| 1476 | // in this function. |
| 1477 | // |
| 1478 | // NOTE: It doesn't make sense to promote non-single-value types since we |
| 1479 | // are just replacing static memory to stack memory. |
| 1480 | // |
| 1481 | // If the global is in different address space, don't bring it to stack. |
| 1482 | if (!GS.HasMultipleAccessingFunctions && |
| 1483 | GS.AccessingFunction && |
| 1484 | GV->getValueType()->isSingleValueType() && |
| 1485 | GV->getType()->getAddressSpace() == DL.getAllocaAddrSpace() && |
| 1486 | !GV->isExternallyInitialized() && |
| 1487 | GS.AccessingFunction->doesNotRecurse() && |
| 1488 | isPointerValueDeadOnEntryToFunction(F: GS.AccessingFunction, GV, |
| 1489 | LookupDomTree)) { |
| 1490 | const DataLayout &DL = GV->getDataLayout(); |
| 1491 | |
| 1492 | LLVM_DEBUG(dbgs() << "LOCALIZING GLOBAL: " << *GV << "\n" ); |
| 1493 | BasicBlock::iterator FirstI = |
| 1494 | GS.AccessingFunction->getEntryBlock().begin().getNonConst(); |
| 1495 | Type *ElemTy = GV->getValueType(); |
| 1496 | // FIXME: Pass Global's alignment when globals have alignment |
| 1497 | AllocaInst *Alloca = new AllocaInst(ElemTy, DL.getAllocaAddrSpace(), |
| 1498 | nullptr, GV->getName(), FirstI); |
| 1499 | Alloca->setDebugLoc(DebugLoc::getCompilerGenerated()); |
| 1500 | if (!isa<UndefValue>(Val: GV->getInitializer())) { |
| 1501 | auto *SI = new StoreInst(GV->getInitializer(), Alloca, FirstI); |
| 1502 | // FIXME: We're localizing a global and creating a store instruction for |
| 1503 | // the initial value of that global. Could we logically use the global |
| 1504 | // variable's (if one exists) line for this? |
| 1505 | SI->setDebugLoc(DebugLoc::getCompilerGenerated()); |
| 1506 | } |
| 1507 | |
| 1508 | GV->replaceAllUsesWith(V: Alloca); |
| 1509 | GV->eraseFromParent(); |
| 1510 | ++NumLocalized; |
| 1511 | return true; |
| 1512 | } |
| 1513 | |
| 1514 | bool Changed = false; |
| 1515 | |
| 1516 | // If the global is never loaded (but may be stored to), it is dead. |
| 1517 | // Delete it now. |
| 1518 | if (!GS.IsLoaded) { |
| 1519 | LLVM_DEBUG(dbgs() << "GLOBAL NEVER LOADED: " << *GV << "\n" ); |
| 1520 | |
| 1521 | if (isLeakCheckerRoot(GV)) { |
| 1522 | // Delete any constant stores to the global. |
| 1523 | Changed = CleanupPointerRootUsers(GV, GetTLI); |
| 1524 | } else { |
| 1525 | // Delete any stores we can find to the global. We may not be able to |
| 1526 | // make it completely dead though. |
| 1527 | Changed = CleanupConstantGlobalUsers(GV, DL); |
| 1528 | } |
| 1529 | |
| 1530 | // If the global is dead now, delete it. |
| 1531 | if (GV->use_empty()) { |
| 1532 | GV->eraseFromParent(); |
| 1533 | ++NumDeleted; |
| 1534 | Changed = true; |
| 1535 | } |
| 1536 | return Changed; |
| 1537 | |
| 1538 | } |
| 1539 | if (GS.StoredType <= GlobalStatus::InitializerStored) { |
| 1540 | LLVM_DEBUG(dbgs() << "MARKING CONSTANT: " << *GV << "\n" ); |
| 1541 | |
| 1542 | // Don't actually mark a global constant if it's atomic because atomic loads |
| 1543 | // are implemented by a trivial cmpxchg in some edge-cases and that usually |
| 1544 | // requires write access to the variable even if it's not actually changed. |
| 1545 | if (GS.Ordering == AtomicOrdering::NotAtomic) { |
| 1546 | assert(!GV->isConstant() && "Expected a non-constant global" ); |
| 1547 | GV->setConstant(true); |
| 1548 | Changed = true; |
| 1549 | } |
| 1550 | |
| 1551 | // Clean up any obviously simplifiable users now. |
| 1552 | Changed |= CleanupConstantGlobalUsers(GV, DL); |
| 1553 | |
| 1554 | // If the global is dead now, just nuke it. |
| 1555 | if (GV->use_empty()) { |
| 1556 | LLVM_DEBUG(dbgs() << " *** Marking constant allowed us to simplify " |
| 1557 | << "all users and delete global!\n" ); |
| 1558 | GV->eraseFromParent(); |
| 1559 | ++NumDeleted; |
| 1560 | return true; |
| 1561 | } |
| 1562 | |
| 1563 | // Fall through to the next check; see if we can optimize further. |
| 1564 | ++NumMarked; |
| 1565 | } |
| 1566 | if (!GV->getInitializer()->getType()->isSingleValueType()) { |
| 1567 | const DataLayout &DL = GV->getDataLayout(); |
| 1568 | if (SRAGlobal(GV, DL)) |
| 1569 | return true; |
| 1570 | } |
| 1571 | Value *StoredOnceValue = GS.getStoredOnceValue(); |
| 1572 | if (GS.StoredType == GlobalStatus::StoredOnce && StoredOnceValue) { |
| 1573 | Function &StoreFn = |
| 1574 | const_cast<Function &>(*GS.StoredOnceStore->getFunction()); |
| 1575 | bool CanHaveNonUndefGlobalInitializer = |
| 1576 | GetTTI(StoreFn).canHaveNonUndefGlobalInitializerInAddressSpace( |
| 1577 | AS: GV->getType()->getAddressSpace()); |
| 1578 | // If the initial value for the global was an undef value, and if only |
| 1579 | // one other value was stored into it, we can just change the |
| 1580 | // initializer to be the stored value, then delete all stores to the |
| 1581 | // global. This allows us to mark it constant. |
| 1582 | // This is restricted to address spaces that allow globals to have |
| 1583 | // initializers. NVPTX, for example, does not support initializers for |
| 1584 | // shared memory (AS 3). |
| 1585 | auto *SOVConstant = dyn_cast<Constant>(Val: StoredOnceValue); |
| 1586 | if (SOVConstant && isa<UndefValue>(Val: GV->getInitializer()) && |
| 1587 | DL.getTypeAllocSize(Ty: SOVConstant->getType()).getFixedValue() == |
| 1588 | GV->getGlobalSize(DL) && |
| 1589 | CanHaveNonUndefGlobalInitializer) { |
| 1590 | if (SOVConstant->getType() == GV->getValueType()) { |
| 1591 | // Change the initializer in place. |
| 1592 | GV->setInitializer(SOVConstant); |
| 1593 | } else { |
| 1594 | // Create a new global with adjusted type. |
| 1595 | auto *NGV = new GlobalVariable( |
| 1596 | *GV->getParent(), SOVConstant->getType(), GV->isConstant(), |
| 1597 | GV->getLinkage(), SOVConstant, "" , GV, GV->getThreadLocalMode(), |
| 1598 | GV->getAddressSpace()); |
| 1599 | NGV->takeName(V: GV); |
| 1600 | NGV->copyAttributesFrom(Src: GV); |
| 1601 | GV->replaceAllUsesWith(V: NGV); |
| 1602 | GV->eraseFromParent(); |
| 1603 | GV = NGV; |
| 1604 | } |
| 1605 | |
| 1606 | // Clean up any obviously simplifiable users now. |
| 1607 | CleanupConstantGlobalUsers(GV, DL); |
| 1608 | |
| 1609 | if (GV->use_empty()) { |
| 1610 | LLVM_DEBUG(dbgs() << " *** Substituting initializer allowed us to " |
| 1611 | << "simplify all users and delete global!\n" ); |
| 1612 | GV->eraseFromParent(); |
| 1613 | ++NumDeleted; |
| 1614 | } |
| 1615 | ++NumSubstitute; |
| 1616 | return true; |
| 1617 | } |
| 1618 | |
| 1619 | // Try to optimize globals based on the knowledge that only one value |
| 1620 | // (besides its initializer) is ever stored to the global. |
| 1621 | if (optimizeOnceStoredGlobal(GV, StoredOnceVal: StoredOnceValue, DL, GetTLI)) |
| 1622 | return true; |
| 1623 | |
| 1624 | // Try to forward the store to any loads. If we have more than one store, we |
| 1625 | // may have a store of the initializer between StoredOnceStore and a load. |
| 1626 | if (GS.NumStores == 1) |
| 1627 | if (forwardStoredOnceStore(GV, StoredOnceStore: GS.StoredOnceStore, LookupDomTree)) |
| 1628 | return true; |
| 1629 | |
| 1630 | // Otherwise, if the global was not a boolean, we can shrink it to be a |
| 1631 | // boolean. Skip this optimization for AS that doesn't allow an initializer. |
| 1632 | if (SOVConstant && GS.Ordering == AtomicOrdering::NotAtomic && |
| 1633 | (!isa<UndefValue>(Val: GV->getInitializer()) || |
| 1634 | CanHaveNonUndefGlobalInitializer)) { |
| 1635 | if (TryToShrinkGlobalToBoolean(GV, OtherVal: SOVConstant)) { |
| 1636 | ++NumShrunkToBool; |
| 1637 | return true; |
| 1638 | } |
| 1639 | } |
| 1640 | } |
| 1641 | |
| 1642 | return Changed; |
| 1643 | } |
| 1644 | |
| 1645 | /// Analyze the specified global variable and optimize it if possible. If we |
| 1646 | /// make a change, return true. |
| 1647 | static bool |
| 1648 | processGlobal(GlobalValue &GV, |
| 1649 | function_ref<TargetTransformInfo &(Function &)> GetTTI, |
| 1650 | function_ref<TargetLibraryInfo &(Function &)> GetTLI, |
| 1651 | function_ref<DominatorTree &(Function &)> LookupDomTree) { |
| 1652 | if (GV.getName().starts_with(Prefix: "llvm." )) |
| 1653 | return false; |
| 1654 | |
| 1655 | GlobalStatus GS; |
| 1656 | |
| 1657 | if (GlobalStatus::analyzeGlobal(V: &GV, GS)) |
| 1658 | return false; |
| 1659 | |
| 1660 | bool Changed = false; |
| 1661 | if (!GS.IsCompared && !GV.hasGlobalUnnamedAddr()) { |
| 1662 | auto NewUnnamedAddr = GV.hasLocalLinkage() ? GlobalValue::UnnamedAddr::Global |
| 1663 | : GlobalValue::UnnamedAddr::Local; |
| 1664 | if (NewUnnamedAddr != GV.getUnnamedAddr()) { |
| 1665 | GV.setUnnamedAddr(NewUnnamedAddr); |
| 1666 | NumUnnamed++; |
| 1667 | Changed = true; |
| 1668 | } |
| 1669 | } |
| 1670 | |
| 1671 | // Do more involved optimizations if the global is internal. |
| 1672 | if (!GV.hasLocalLinkage()) |
| 1673 | return Changed; |
| 1674 | |
| 1675 | auto *GVar = dyn_cast<GlobalVariable>(Val: &GV); |
| 1676 | if (!GVar) |
| 1677 | return Changed; |
| 1678 | |
| 1679 | if (GVar->isConstant() || !GVar->hasInitializer()) |
| 1680 | return Changed; |
| 1681 | |
| 1682 | return processInternalGlobal(GV: GVar, GS, GetTTI, GetTLI, LookupDomTree) || |
| 1683 | Changed; |
| 1684 | } |
| 1685 | |
| 1686 | /// Walk all of the direct calls of the specified function, changing them to |
| 1687 | /// FastCC. |
| 1688 | static void ChangeCalleesToFastCall(Function *F) { |
| 1689 | for (User *U : F->users()) |
| 1690 | if (auto *Call = dyn_cast<CallBase>(Val: U)) |
| 1691 | if (Call->getCalledOperand() == F) |
| 1692 | Call->setCallingConv(CallingConv::Fast); |
| 1693 | } |
| 1694 | |
| 1695 | static AttributeList StripAttr(LLVMContext &C, AttributeList Attrs, |
| 1696 | Attribute::AttrKind A) { |
| 1697 | unsigned AttrIndex; |
| 1698 | if (Attrs.hasAttrSomewhere(Kind: A, Index: &AttrIndex)) |
| 1699 | return Attrs.removeAttributeAtIndex(C, Index: AttrIndex, Kind: A); |
| 1700 | return Attrs; |
| 1701 | } |
| 1702 | |
| 1703 | static void RemoveAttribute(Function *F, Attribute::AttrKind A) { |
| 1704 | F->setAttributes(StripAttr(C&: F->getContext(), Attrs: F->getAttributes(), A)); |
| 1705 | for (User *U : F->users()) { |
| 1706 | CallBase *CB = cast<CallBase>(Val: U); |
| 1707 | CB->setAttributes(StripAttr(C&: F->getContext(), Attrs: CB->getAttributes(), A)); |
| 1708 | } |
| 1709 | } |
| 1710 | |
| 1711 | /// Return true if this is a calling convention that we'd like to change. The |
| 1712 | /// idea here is that we don't want to mess with the convention if the user |
| 1713 | /// explicitly requested something with performance implications like coldcc, |
| 1714 | /// GHC, or anyregcc. |
| 1715 | static bool hasChangeableCCImpl(Function *F) { |
| 1716 | CallingConv::ID CC = F->getCallingConv(); |
| 1717 | |
| 1718 | // FIXME: Is it worth transforming x86_stdcallcc and x86_fastcallcc? |
| 1719 | if (CC != CallingConv::C && CC != CallingConv::X86_ThisCall) |
| 1720 | return false; |
| 1721 | |
| 1722 | if (F->isVarArg()) |
| 1723 | return false; |
| 1724 | |
| 1725 | // FIXME: Change CC for the whole chain of musttail calls when possible. |
| 1726 | // |
| 1727 | // Can't change CC of the function that either has musttail calls, or is a |
| 1728 | // musttail callee itself |
| 1729 | for (User *U : F->users()) { |
| 1730 | CallInst* CI = dyn_cast<CallInst>(Val: U); |
| 1731 | if (!CI) |
| 1732 | continue; |
| 1733 | |
| 1734 | if (CI->isMustTailCall()) |
| 1735 | return false; |
| 1736 | } |
| 1737 | |
| 1738 | for (BasicBlock &BB : *F) |
| 1739 | if (BB.getTerminatingMustTailCall()) |
| 1740 | return false; |
| 1741 | |
| 1742 | return !F->hasAddressTaken(); |
| 1743 | } |
| 1744 | |
| 1745 | using ChangeableCCCacheTy = SmallDenseMap<Function *, bool, 8>; |
| 1746 | static bool hasChangeableCC(Function *F, |
| 1747 | ChangeableCCCacheTy &ChangeableCCCache) { |
| 1748 | auto Res = ChangeableCCCache.try_emplace(Key: F, Args: false); |
| 1749 | if (Res.second) |
| 1750 | Res.first->second = hasChangeableCCImpl(F); |
| 1751 | return Res.first->second; |
| 1752 | } |
| 1753 | |
| 1754 | /// Return true if the block containing the call site has a BlockFrequency of |
| 1755 | /// less than ColdCCRelFreq% of the entry block. |
| 1756 | static bool isColdCallSite(CallBase &CB, BlockFrequencyInfo &CallerBFI) { |
| 1757 | const BranchProbability ColdProb(ColdCCRelFreq, 100); |
| 1758 | auto *CallSiteBB = CB.getParent(); |
| 1759 | auto CallSiteFreq = CallerBFI.getBlockFreq(BB: CallSiteBB); |
| 1760 | auto CallerEntryFreq = |
| 1761 | CallerBFI.getBlockFreq(BB: &(CB.getCaller()->getEntryBlock())); |
| 1762 | return CallSiteFreq < CallerEntryFreq * ColdProb; |
| 1763 | } |
| 1764 | |
| 1765 | // This function checks if the input function F is cold at all call sites. It |
| 1766 | // also looks each call site's containing function, returning false if the |
| 1767 | // caller function contains other non cold calls. The input vector AllCallsCold |
| 1768 | // contains a list of functions that only have call sites in cold blocks. |
| 1769 | static bool |
| 1770 | isValidCandidateForColdCC(Function &F, |
| 1771 | function_ref<BlockFrequencyInfo &(Function &)> GetBFI, |
| 1772 | const std::vector<Function *> &AllCallsCold) { |
| 1773 | |
| 1774 | if (F.user_empty()) |
| 1775 | return false; |
| 1776 | |
| 1777 | for (User *U : F.users()) { |
| 1778 | CallBase *CB = dyn_cast<CallBase>(Val: U); |
| 1779 | if (!CB || CB->getCalledOperand() != &F) |
| 1780 | continue; |
| 1781 | Function *CallerFunc = CB->getParent()->getParent(); |
| 1782 | BlockFrequencyInfo &CallerBFI = GetBFI(*CallerFunc); |
| 1783 | if (!isColdCallSite(CB&: *CB, CallerBFI)) |
| 1784 | return false; |
| 1785 | if (!llvm::is_contained(Range: AllCallsCold, Element: CallerFunc)) |
| 1786 | return false; |
| 1787 | } |
| 1788 | return true; |
| 1789 | } |
| 1790 | |
| 1791 | static void changeCallSitesToColdCC(Function *F) { |
| 1792 | for (User *U : F->users()) |
| 1793 | if (auto *Call = dyn_cast<CallBase>(Val: U)) |
| 1794 | if (Call->getCalledOperand() == F) |
| 1795 | Call->setCallingConv(CallingConv::Cold); |
| 1796 | } |
| 1797 | |
| 1798 | // This function iterates over all the call instructions in the input Function |
| 1799 | // and checks that all call sites are in cold blocks and are allowed to use the |
| 1800 | // coldcc calling convention. |
| 1801 | static bool |
| 1802 | hasOnlyColdCalls(Function &F, |
| 1803 | function_ref<BlockFrequencyInfo &(Function &)> GetBFI, |
| 1804 | ChangeableCCCacheTy &ChangeableCCCache) { |
| 1805 | for (BasicBlock &BB : F) { |
| 1806 | for (Instruction &I : BB) { |
| 1807 | if (CallInst *CI = dyn_cast<CallInst>(Val: &I)) { |
| 1808 | // Skip over isline asm instructions since they aren't function calls. |
| 1809 | if (CI->isInlineAsm()) |
| 1810 | continue; |
| 1811 | Function *CalledFn = CI->getCalledFunction(); |
| 1812 | if (!CalledFn) |
| 1813 | return false; |
| 1814 | // Skip over intrinsics since they won't remain as function calls. |
| 1815 | // Important to do this check before the linkage check below so we |
| 1816 | // won't bail out on debug intrinsics, possibly making the generated |
| 1817 | // code dependent on the presence of debug info. |
| 1818 | if (CalledFn->getIntrinsicID() != Intrinsic::not_intrinsic) |
| 1819 | continue; |
| 1820 | if (!CalledFn->hasLocalLinkage()) |
| 1821 | return false; |
| 1822 | // Check if it's valid to use coldcc calling convention. |
| 1823 | if (!hasChangeableCC(F: CalledFn, ChangeableCCCache)) |
| 1824 | return false; |
| 1825 | BlockFrequencyInfo &CallerBFI = GetBFI(F); |
| 1826 | if (!isColdCallSite(CB&: *CI, CallerBFI)) |
| 1827 | return false; |
| 1828 | } |
| 1829 | } |
| 1830 | } |
| 1831 | return true; |
| 1832 | } |
| 1833 | |
| 1834 | static bool hasMustTailCallers(Function *F) { |
| 1835 | for (User *U : F->users()) { |
| 1836 | CallBase *CB = cast<CallBase>(Val: U); |
| 1837 | if (CB->isMustTailCall()) |
| 1838 | return true; |
| 1839 | } |
| 1840 | return false; |
| 1841 | } |
| 1842 | |
| 1843 | static bool hasInvokeCallers(Function *F) { |
| 1844 | for (User *U : F->users()) |
| 1845 | if (isa<InvokeInst>(Val: U)) |
| 1846 | return true; |
| 1847 | return false; |
| 1848 | } |
| 1849 | |
| 1850 | static void RemovePreallocated(Function *F) { |
| 1851 | RemoveAttribute(F, A: Attribute::Preallocated); |
| 1852 | |
| 1853 | auto *M = F->getParent(); |
| 1854 | |
| 1855 | IRBuilder<> Builder(M->getContext()); |
| 1856 | |
| 1857 | // Cannot modify users() while iterating over it, so make a copy. |
| 1858 | SmallVector<User *, 4> PreallocatedCalls(F->users()); |
| 1859 | for (User *U : PreallocatedCalls) { |
| 1860 | CallBase *CB = dyn_cast<CallBase>(Val: U); |
| 1861 | if (!CB) |
| 1862 | continue; |
| 1863 | |
| 1864 | assert( |
| 1865 | !CB->isMustTailCall() && |
| 1866 | "Shouldn't call RemotePreallocated() on a musttail preallocated call" ); |
| 1867 | // Create copy of call without "preallocated" operand bundle. |
| 1868 | SmallVector<OperandBundleDef, 1> OpBundles; |
| 1869 | CB->getOperandBundlesAsDefs(Defs&: OpBundles); |
| 1870 | CallBase *PreallocatedSetup = nullptr; |
| 1871 | for (auto *It = OpBundles.begin(); It != OpBundles.end(); ++It) { |
| 1872 | if (It->getTag() == "preallocated" ) { |
| 1873 | PreallocatedSetup = cast<CallBase>(Val: *It->input_begin()); |
| 1874 | OpBundles.erase(CI: It); |
| 1875 | break; |
| 1876 | } |
| 1877 | } |
| 1878 | assert(PreallocatedSetup && "Did not find preallocated bundle" ); |
| 1879 | uint64_t ArgCount = |
| 1880 | cast<ConstantInt>(Val: PreallocatedSetup->getArgOperand(i: 0))->getZExtValue(); |
| 1881 | |
| 1882 | assert((isa<CallInst>(CB) || isa<InvokeInst>(CB)) && |
| 1883 | "Unknown indirect call type" ); |
| 1884 | CallBase *NewCB = CallBase::Create(CB, Bundles: OpBundles, InsertPt: CB->getIterator()); |
| 1885 | CB->replaceAllUsesWith(V: NewCB); |
| 1886 | NewCB->takeName(V: CB); |
| 1887 | CB->eraseFromParent(); |
| 1888 | |
| 1889 | Builder.SetInsertPoint(PreallocatedSetup); |
| 1890 | auto *StackSave = Builder.CreateStackSave(); |
| 1891 | Builder.SetInsertPoint(NewCB->getNextNode()); |
| 1892 | Builder.CreateStackRestore(Ptr: StackSave); |
| 1893 | |
| 1894 | // Replace @llvm.call.preallocated.arg() with alloca. |
| 1895 | // Cannot modify users() while iterating over it, so make a copy. |
| 1896 | // @llvm.call.preallocated.arg() can be called with the same index multiple |
| 1897 | // times. So for each @llvm.call.preallocated.arg(), we see if we have |
| 1898 | // already created a Value* for the index, and if not, create an alloca and |
| 1899 | // bitcast right after the @llvm.call.preallocated.setup() so that it |
| 1900 | // dominates all uses. |
| 1901 | SmallVector<Value *, 2> ArgAllocas(ArgCount); |
| 1902 | SmallVector<User *, 2> PreallocatedArgs(PreallocatedSetup->users()); |
| 1903 | for (auto *User : PreallocatedArgs) { |
| 1904 | auto *UseCall = cast<CallBase>(Val: User); |
| 1905 | assert(UseCall->getCalledFunction()->getIntrinsicID() == |
| 1906 | Intrinsic::call_preallocated_arg && |
| 1907 | "preallocated token use was not a llvm.call.preallocated.arg" ); |
| 1908 | uint64_t AllocArgIndex = |
| 1909 | cast<ConstantInt>(Val: UseCall->getArgOperand(i: 1))->getZExtValue(); |
| 1910 | Value *AllocaReplacement = ArgAllocas[AllocArgIndex]; |
| 1911 | if (!AllocaReplacement) { |
| 1912 | auto AddressSpace = UseCall->getType()->getPointerAddressSpace(); |
| 1913 | auto *ArgType = |
| 1914 | UseCall->getFnAttr(Kind: Attribute::Preallocated).getValueAsType(); |
| 1915 | auto *InsertBefore = PreallocatedSetup->getNextNode(); |
| 1916 | Builder.SetInsertPoint(InsertBefore); |
| 1917 | auto *Alloca = |
| 1918 | Builder.CreateAlloca(Ty: ArgType, AddrSpace: AddressSpace, ArraySize: nullptr, Name: "paarg" ); |
| 1919 | ArgAllocas[AllocArgIndex] = Alloca; |
| 1920 | AllocaReplacement = Alloca; |
| 1921 | } |
| 1922 | |
| 1923 | UseCall->replaceAllUsesWith(V: AllocaReplacement); |
| 1924 | UseCall->eraseFromParent(); |
| 1925 | } |
| 1926 | // Remove @llvm.call.preallocated.setup(). |
| 1927 | cast<Instruction>(Val: PreallocatedSetup)->eraseFromParent(); |
| 1928 | } |
| 1929 | } |
| 1930 | |
| 1931 | static bool |
| 1932 | OptimizeFunctions(Module &M, |
| 1933 | function_ref<TargetLibraryInfo &(Function &)> GetTLI, |
| 1934 | function_ref<TargetTransformInfo &(Function &)> GetTTI, |
| 1935 | function_ref<BlockFrequencyInfo &(Function &)> GetBFI, |
| 1936 | function_ref<DominatorTree &(Function &)> LookupDomTree, |
| 1937 | SmallPtrSetImpl<const Comdat *> &NotDiscardableComdats, |
| 1938 | function_ref<void(Function &F)> ChangedCFGCallback, |
| 1939 | function_ref<void(Function &F)> DeleteFnCallback) { |
| 1940 | |
| 1941 | bool Changed = false; |
| 1942 | |
| 1943 | ChangeableCCCacheTy ChangeableCCCache; |
| 1944 | std::vector<Function *> AllCallsCold; |
| 1945 | for (Function &F : llvm::make_early_inc_range(Range&: M)) |
| 1946 | if (hasOnlyColdCalls(F, GetBFI, ChangeableCCCache)) |
| 1947 | AllCallsCold.push_back(x: &F); |
| 1948 | |
| 1949 | // Optimize functions. |
| 1950 | for (Function &F : llvm::make_early_inc_range(Range&: M)) { |
| 1951 | // Don't perform global opt pass on naked functions; we don't want fast |
| 1952 | // calling conventions for naked functions. |
| 1953 | if (F.hasFnAttribute(Kind: Attribute::Naked)) |
| 1954 | continue; |
| 1955 | |
| 1956 | // Functions without names cannot be referenced outside this module. |
| 1957 | if (!F.hasName() && !F.isDeclaration() && !F.hasLocalLinkage()) |
| 1958 | F.setLinkage(GlobalValue::InternalLinkage); |
| 1959 | |
| 1960 | if (deleteIfDead(GV&: F, NotDiscardableComdats, DeleteFnCallback)) { |
| 1961 | Changed = true; |
| 1962 | continue; |
| 1963 | } |
| 1964 | |
| 1965 | // LLVM's definition of dominance allows instructions that are cyclic |
| 1966 | // in unreachable blocks, e.g.: |
| 1967 | // %pat = select i1 %condition, @global, i16* %pat |
| 1968 | // because any instruction dominates an instruction in a block that's |
| 1969 | // not reachable from entry. |
| 1970 | // So, remove unreachable blocks from the function, because a) there's |
| 1971 | // no point in analyzing them and b) GlobalOpt should otherwise grow |
| 1972 | // some more complicated logic to break these cycles. |
| 1973 | // Notify the analysis manager that we've modified the function's CFG. |
| 1974 | if (!F.isDeclaration()) { |
| 1975 | if (removeUnreachableBlocks(F)) { |
| 1976 | Changed = true; |
| 1977 | ChangedCFGCallback(F); |
| 1978 | } |
| 1979 | } |
| 1980 | |
| 1981 | Changed |= processGlobal(GV&: F, GetTTI, GetTLI, LookupDomTree); |
| 1982 | |
| 1983 | if (!F.hasLocalLinkage()) |
| 1984 | continue; |
| 1985 | |
| 1986 | // If we have an inalloca parameter that we can safely remove the |
| 1987 | // inalloca attribute from, do so. This unlocks optimizations that |
| 1988 | // wouldn't be safe in the presence of inalloca. |
| 1989 | // FIXME: We should also hoist alloca affected by this to the entry |
| 1990 | // block if possible. |
| 1991 | if (F.getAttributes().hasAttrSomewhere(Kind: Attribute::InAlloca) && |
| 1992 | !F.hasAddressTaken() && !hasMustTailCallers(F: &F) && !F.isVarArg()) { |
| 1993 | RemoveAttribute(F: &F, A: Attribute::InAlloca); |
| 1994 | Changed = true; |
| 1995 | } |
| 1996 | |
| 1997 | // FIXME: handle invokes |
| 1998 | // FIXME: handle musttail |
| 1999 | if (F.getAttributes().hasAttrSomewhere(Kind: Attribute::Preallocated)) { |
| 2000 | if (!F.hasAddressTaken() && !hasMustTailCallers(F: &F) && |
| 2001 | !hasInvokeCallers(F: &F)) { |
| 2002 | RemovePreallocated(F: &F); |
| 2003 | Changed = true; |
| 2004 | } |
| 2005 | continue; |
| 2006 | } |
| 2007 | |
| 2008 | if (hasChangeableCC(F: &F, ChangeableCCCache)) { |
| 2009 | NumInternalFunc++; |
| 2010 | TargetTransformInfo &TTI = GetTTI(F); |
| 2011 | // Change the calling convention to coldcc if either stress testing is |
| 2012 | // enabled or the target would like to use coldcc on functions which are |
| 2013 | // cold at all call sites and the callers contain no other non coldcc |
| 2014 | // calls. |
| 2015 | if (EnableColdCCStressTest || |
| 2016 | (TTI.useColdCCForColdCall(F) && |
| 2017 | isValidCandidateForColdCC(F, GetBFI, AllCallsCold))) { |
| 2018 | ChangeableCCCache.erase(Val: &F); |
| 2019 | F.setCallingConv(CallingConv::Cold); |
| 2020 | changeCallSitesToColdCC(F: &F); |
| 2021 | Changed = true; |
| 2022 | NumColdCC++; |
| 2023 | } |
| 2024 | } |
| 2025 | |
| 2026 | if (hasChangeableCC(F: &F, ChangeableCCCache)) { |
| 2027 | // If this function has a calling convention worth changing, is not a |
| 2028 | // varargs function, is only called directly, and is supported by the |
| 2029 | // target, promote it to use the Fast calling convention. |
| 2030 | TargetTransformInfo &TTI = GetTTI(F); |
| 2031 | if (TTI.useFastCCForInternalCall(F)) { |
| 2032 | F.setCallingConv(CallingConv::Fast); |
| 2033 | ChangeCalleesToFastCall(F: &F); |
| 2034 | ++NumFastCallFns; |
| 2035 | Changed = true; |
| 2036 | } |
| 2037 | } |
| 2038 | |
| 2039 | if (F.getAttributes().hasAttrSomewhere(Kind: Attribute::Nest) && |
| 2040 | !F.hasAddressTaken()) { |
| 2041 | // The function is not used by a trampoline intrinsic, so it is safe |
| 2042 | // to remove the 'nest' attribute. |
| 2043 | RemoveAttribute(F: &F, A: Attribute::Nest); |
| 2044 | ++NumNestRemoved; |
| 2045 | Changed = true; |
| 2046 | } |
| 2047 | } |
| 2048 | return Changed; |
| 2049 | } |
| 2050 | |
| 2051 | static bool |
| 2052 | OptimizeGlobalVars(Module &M, |
| 2053 | function_ref<TargetTransformInfo &(Function &)> GetTTI, |
| 2054 | function_ref<TargetLibraryInfo &(Function &)> GetTLI, |
| 2055 | function_ref<DominatorTree &(Function &)> LookupDomTree, |
| 2056 | SmallPtrSetImpl<const Comdat *> &NotDiscardableComdats) { |
| 2057 | bool Changed = false; |
| 2058 | |
| 2059 | for (GlobalVariable &GV : llvm::make_early_inc_range(Range: M.globals())) { |
| 2060 | // Global variables without names cannot be referenced outside this module. |
| 2061 | if (!GV.hasName() && !GV.isDeclaration() && !GV.hasLocalLinkage()) |
| 2062 | GV.setLinkage(GlobalValue::InternalLinkage); |
| 2063 | // Simplify the initializer. |
| 2064 | if (GV.hasInitializer()) |
| 2065 | if (auto *C = dyn_cast<Constant>(Val: GV.getInitializer())) { |
| 2066 | auto &DL = M.getDataLayout(); |
| 2067 | // TLI is not used in the case of a Constant, so use default nullptr |
| 2068 | // for that optional parameter, since we don't have a Function to |
| 2069 | // provide GetTLI anyway. |
| 2070 | Constant *New = ConstantFoldConstant(C, DL, /*TLI*/ nullptr); |
| 2071 | if (New != C) |
| 2072 | GV.setInitializer(New); |
| 2073 | } |
| 2074 | |
| 2075 | if (deleteIfDead(GV, NotDiscardableComdats)) { |
| 2076 | Changed = true; |
| 2077 | continue; |
| 2078 | } |
| 2079 | |
| 2080 | Changed |= processGlobal(GV, GetTTI, GetTLI, LookupDomTree); |
| 2081 | } |
| 2082 | return Changed; |
| 2083 | } |
| 2084 | |
| 2085 | /// Evaluate static constructors in the function, if we can. Return true if we |
| 2086 | /// can, false otherwise. |
| 2087 | static bool EvaluateStaticConstructor(Function *F, const DataLayout &DL, |
| 2088 | TargetLibraryInfo *TLI) { |
| 2089 | // Skip external functions. |
| 2090 | if (F->isDeclaration()) |
| 2091 | return false; |
| 2092 | // Call the function. |
| 2093 | Evaluator Eval(DL, TLI); |
| 2094 | Constant *RetValDummy; |
| 2095 | bool EvalSuccess = Eval.EvaluateFunction(F, RetVal&: RetValDummy, |
| 2096 | ActualArgs: SmallVector<Constant*, 0>()); |
| 2097 | |
| 2098 | if (EvalSuccess) { |
| 2099 | ++NumCtorsEvaluated; |
| 2100 | |
| 2101 | // We succeeded at evaluation: commit the result. |
| 2102 | auto NewInitializers = Eval.getMutatedInitializers(); |
| 2103 | LLVM_DEBUG(dbgs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '" |
| 2104 | << F->getName() << "' to " << NewInitializers.size() |
| 2105 | << " stores.\n" ); |
| 2106 | for (const auto &Pair : NewInitializers) |
| 2107 | Pair.first->setInitializer(Pair.second); |
| 2108 | for (GlobalVariable *GV : Eval.getInvariants()) |
| 2109 | GV->setConstant(true); |
| 2110 | } |
| 2111 | |
| 2112 | return EvalSuccess; |
| 2113 | } |
| 2114 | |
| 2115 | static int compareNames(Constant *const *A, Constant *const *B) { |
| 2116 | Value *AStripped = (*A)->stripPointerCasts(); |
| 2117 | Value *BStripped = (*B)->stripPointerCasts(); |
| 2118 | return AStripped->getName().compare(RHS: BStripped->getName()); |
| 2119 | } |
| 2120 | |
| 2121 | static void setUsedInitializer(GlobalVariable &V, |
| 2122 | const SmallPtrSetImpl<GlobalValue *> &Init) { |
| 2123 | if (Init.empty()) { |
| 2124 | V.eraseFromParent(); |
| 2125 | return; |
| 2126 | } |
| 2127 | |
| 2128 | // Get address space of pointers in the array of pointers. |
| 2129 | const Type *UsedArrayType = V.getValueType(); |
| 2130 | const auto *VAT = cast<ArrayType>(Val: UsedArrayType); |
| 2131 | const auto *VEPT = cast<PointerType>(Val: VAT->getArrayElementType()); |
| 2132 | |
| 2133 | // Type of pointer to the array of pointers. |
| 2134 | PointerType *PtrTy = |
| 2135 | PointerType::get(C&: V.getContext(), AddressSpace: VEPT->getAddressSpace()); |
| 2136 | |
| 2137 | SmallVector<Constant *, 8> UsedArray; |
| 2138 | for (GlobalValue *GV : Init) { |
| 2139 | Constant *Cast = ConstantExpr::getPointerBitCastOrAddrSpaceCast(C: GV, Ty: PtrTy); |
| 2140 | UsedArray.push_back(Elt: Cast); |
| 2141 | } |
| 2142 | |
| 2143 | // Sort to get deterministic order. |
| 2144 | array_pod_sort(Start: UsedArray.begin(), End: UsedArray.end(), Compare: compareNames); |
| 2145 | ArrayType *ATy = ArrayType::get(ElementType: PtrTy, NumElements: UsedArray.size()); |
| 2146 | |
| 2147 | Module *M = V.getParent(); |
| 2148 | V.removeFromParent(); |
| 2149 | GlobalVariable *NV = new GlobalVariable( |
| 2150 | *M, ATy, false, GlobalValue::AppendingLinkage, |
| 2151 | ConstantArray::get(T: ATy, V: UsedArray), "" , nullptr, |
| 2152 | GlobalVariable::NotThreadLocal, V.getType()->getAddressSpace()); |
| 2153 | NV->takeName(V: &V); |
| 2154 | NV->setSection("llvm.metadata" ); |
| 2155 | delete &V; |
| 2156 | } |
| 2157 | |
| 2158 | namespace { |
| 2159 | |
| 2160 | /// An easy to access representation of llvm.used and llvm.compiler.used. |
| 2161 | class LLVMUsed { |
| 2162 | SmallPtrSet<GlobalValue *, 4> Used; |
| 2163 | SmallPtrSet<GlobalValue *, 4> CompilerUsed; |
| 2164 | GlobalVariable *UsedV; |
| 2165 | GlobalVariable *CompilerUsedV; |
| 2166 | |
| 2167 | public: |
| 2168 | LLVMUsed(Module &M) { |
| 2169 | SmallVector<GlobalValue *, 4> Vec; |
| 2170 | UsedV = collectUsedGlobalVariables(M, Vec, CompilerUsed: false); |
| 2171 | Used = {llvm::from_range, Vec}; |
| 2172 | Vec.clear(); |
| 2173 | CompilerUsedV = collectUsedGlobalVariables(M, Vec, CompilerUsed: true); |
| 2174 | CompilerUsed = {llvm::from_range, Vec}; |
| 2175 | } |
| 2176 | |
| 2177 | using iterator = SmallPtrSet<GlobalValue *, 4>::iterator; |
| 2178 | using used_iterator_range = iterator_range<iterator>; |
| 2179 | |
| 2180 | iterator usedBegin() { return Used.begin(); } |
| 2181 | iterator usedEnd() { return Used.end(); } |
| 2182 | |
| 2183 | used_iterator_range used() { |
| 2184 | return used_iterator_range(usedBegin(), usedEnd()); |
| 2185 | } |
| 2186 | |
| 2187 | iterator compilerUsedBegin() { return CompilerUsed.begin(); } |
| 2188 | iterator compilerUsedEnd() { return CompilerUsed.end(); } |
| 2189 | |
| 2190 | used_iterator_range compilerUsed() { |
| 2191 | return used_iterator_range(compilerUsedBegin(), compilerUsedEnd()); |
| 2192 | } |
| 2193 | |
| 2194 | bool usedCount(GlobalValue *GV) const { return Used.count(Ptr: GV); } |
| 2195 | |
| 2196 | bool compilerUsedCount(GlobalValue *GV) const { |
| 2197 | return CompilerUsed.count(Ptr: GV); |
| 2198 | } |
| 2199 | |
| 2200 | bool usedErase(GlobalValue *GV) { return Used.erase(Ptr: GV); } |
| 2201 | bool compilerUsedErase(GlobalValue *GV) { return CompilerUsed.erase(Ptr: GV); } |
| 2202 | bool usedInsert(GlobalValue *GV) { return Used.insert(Ptr: GV).second; } |
| 2203 | |
| 2204 | bool compilerUsedInsert(GlobalValue *GV) { |
| 2205 | return CompilerUsed.insert(Ptr: GV).second; |
| 2206 | } |
| 2207 | |
| 2208 | void syncVariablesAndSets() { |
| 2209 | if (UsedV) |
| 2210 | setUsedInitializer(V&: *UsedV, Init: Used); |
| 2211 | if (CompilerUsedV) |
| 2212 | setUsedInitializer(V&: *CompilerUsedV, Init: CompilerUsed); |
| 2213 | } |
| 2214 | }; |
| 2215 | |
| 2216 | } // end anonymous namespace |
| 2217 | |
| 2218 | static bool hasUseOtherThanLLVMUsed(GlobalAlias &GA, const LLVMUsed &U) { |
| 2219 | if (GA.use_empty()) // No use at all. |
| 2220 | return false; |
| 2221 | |
| 2222 | assert((!U.usedCount(&GA) || !U.compilerUsedCount(&GA)) && |
| 2223 | "We should have removed the duplicated " |
| 2224 | "element from llvm.compiler.used" ); |
| 2225 | if (!GA.hasOneUse()) |
| 2226 | // Strictly more than one use. So at least one is not in llvm.used and |
| 2227 | // llvm.compiler.used. |
| 2228 | return true; |
| 2229 | |
| 2230 | // Exactly one use. Check if it is in llvm.used or llvm.compiler.used. |
| 2231 | return !U.usedCount(GV: &GA) && !U.compilerUsedCount(GV: &GA); |
| 2232 | } |
| 2233 | |
| 2234 | static bool mayHaveOtherReferences(GlobalValue &GV, const LLVMUsed &U) { |
| 2235 | if (!GV.hasLocalLinkage()) |
| 2236 | return true; |
| 2237 | |
| 2238 | return U.usedCount(GV: &GV) || U.compilerUsedCount(GV: &GV); |
| 2239 | } |
| 2240 | |
| 2241 | static bool hasUsesToReplace(GlobalAlias &GA, const LLVMUsed &U, |
| 2242 | bool &RenameTarget) { |
| 2243 | if (GA.isWeakForLinker()) |
| 2244 | return false; |
| 2245 | |
| 2246 | RenameTarget = false; |
| 2247 | bool Ret = false; |
| 2248 | if (hasUseOtherThanLLVMUsed(GA, U)) |
| 2249 | Ret = true; |
| 2250 | |
| 2251 | // If the alias is externally visible, we may still be able to simplify it. |
| 2252 | if (!mayHaveOtherReferences(GV&: GA, U)) |
| 2253 | return Ret; |
| 2254 | |
| 2255 | // If the aliasee has internal linkage and no other references (e.g., |
| 2256 | // @llvm.used, @llvm.compiler.used), give it the name and linkage of the |
| 2257 | // alias, and delete the alias. This turns: |
| 2258 | // define internal ... @f(...) |
| 2259 | // @a = alias ... @f |
| 2260 | // into: |
| 2261 | // define ... @a(...) |
| 2262 | Constant *Aliasee = GA.getAliasee(); |
| 2263 | GlobalValue *Target = cast<GlobalValue>(Val: Aliasee->stripPointerCasts()); |
| 2264 | if (mayHaveOtherReferences(GV&: *Target, U)) |
| 2265 | return Ret; |
| 2266 | |
| 2267 | RenameTarget = true; |
| 2268 | return true; |
| 2269 | } |
| 2270 | |
| 2271 | static bool |
| 2272 | OptimizeGlobalAliases(Module &M, |
| 2273 | SmallPtrSetImpl<const Comdat *> &NotDiscardableComdats) { |
| 2274 | bool Changed = false; |
| 2275 | LLVMUsed Used(M); |
| 2276 | |
| 2277 | for (GlobalValue *GV : Used.used()) |
| 2278 | Used.compilerUsedErase(GV); |
| 2279 | |
| 2280 | // Return whether GV is explicitly or implicitly dso_local and not replaceable |
| 2281 | // by another definition in the current linkage unit. |
| 2282 | auto IsModuleLocal = [](GlobalValue &GV) { |
| 2283 | return !GlobalValue::isInterposableLinkage(Linkage: GV.getLinkage()) && |
| 2284 | (GV.isDSOLocal() || GV.isImplicitDSOLocal()); |
| 2285 | }; |
| 2286 | |
| 2287 | for (GlobalAlias &J : llvm::make_early_inc_range(Range: M.aliases())) { |
| 2288 | // Aliases without names cannot be referenced outside this module. |
| 2289 | if (!J.hasName() && !J.isDeclaration() && !J.hasLocalLinkage()) |
| 2290 | J.setLinkage(GlobalValue::InternalLinkage); |
| 2291 | |
| 2292 | if (deleteIfDead(GV&: J, NotDiscardableComdats)) { |
| 2293 | Changed = true; |
| 2294 | continue; |
| 2295 | } |
| 2296 | |
| 2297 | // If the alias can change at link time, nothing can be done - bail out. |
| 2298 | if (!IsModuleLocal(J)) |
| 2299 | continue; |
| 2300 | |
| 2301 | Constant *Aliasee = J.getAliasee(); |
| 2302 | GlobalValue *Target = dyn_cast<GlobalValue>(Val: Aliasee->stripPointerCasts()); |
| 2303 | // We can't trivially replace the alias with the aliasee if the aliasee is |
| 2304 | // non-trivial in some way. We also can't replace the alias with the aliasee |
| 2305 | // if the aliasee may be preemptible at runtime. On ELF, a non-preemptible |
| 2306 | // alias can be used to access the definition as if preemption did not |
| 2307 | // happen. |
| 2308 | // TODO: Try to handle non-zero GEPs of local aliasees. |
| 2309 | if (!Target || !IsModuleLocal(*Target)) |
| 2310 | continue; |
| 2311 | |
| 2312 | Target->removeDeadConstantUsers(); |
| 2313 | |
| 2314 | // Make all users of the alias use the aliasee instead. |
| 2315 | bool RenameTarget; |
| 2316 | if (!hasUsesToReplace(GA&: J, U: Used, RenameTarget)) |
| 2317 | continue; |
| 2318 | |
| 2319 | J.replaceAllUsesWith(V: Aliasee); |
| 2320 | ++NumAliasesResolved; |
| 2321 | Changed = true; |
| 2322 | |
| 2323 | if (RenameTarget) { |
| 2324 | // Give the aliasee the name, linkage and other attributes of the alias. |
| 2325 | Target->takeName(V: &J); |
| 2326 | Target->setLinkage(J.getLinkage()); |
| 2327 | Target->setDSOLocal(J.isDSOLocal()); |
| 2328 | Target->setVisibility(J.getVisibility()); |
| 2329 | Target->setDLLStorageClass(J.getDLLStorageClass()); |
| 2330 | |
| 2331 | if (Used.usedErase(GV: &J)) |
| 2332 | Used.usedInsert(GV: Target); |
| 2333 | |
| 2334 | if (Used.compilerUsedErase(GV: &J)) |
| 2335 | Used.compilerUsedInsert(GV: Target); |
| 2336 | } else if (mayHaveOtherReferences(GV&: J, U: Used)) |
| 2337 | continue; |
| 2338 | |
| 2339 | // Delete the alias. |
| 2340 | M.eraseAlias(Alias: &J); |
| 2341 | ++NumAliasesRemoved; |
| 2342 | Changed = true; |
| 2343 | } |
| 2344 | |
| 2345 | Used.syncVariablesAndSets(); |
| 2346 | |
| 2347 | return Changed; |
| 2348 | } |
| 2349 | |
| 2350 | static Function * |
| 2351 | FindAtExitLibFunc(Module &M, |
| 2352 | function_ref<TargetLibraryInfo &(Function &)> GetTLI, |
| 2353 | LibFunc Func) { |
| 2354 | // Hack to get a default TLI before we have actual Function. |
| 2355 | auto FuncIter = M.begin(); |
| 2356 | if (FuncIter == M.end()) |
| 2357 | return nullptr; |
| 2358 | auto *TLI = &GetTLI(*FuncIter); |
| 2359 | |
| 2360 | if (!TLI->has(F: Func)) |
| 2361 | return nullptr; |
| 2362 | |
| 2363 | Function *Fn = M.getFunction(Name: TLI->getName(F: Func)); |
| 2364 | if (!Fn) |
| 2365 | return nullptr; |
| 2366 | |
| 2367 | // Now get the actual TLI for Fn. |
| 2368 | TLI = &GetTLI(*Fn); |
| 2369 | |
| 2370 | // Make sure that the function has the correct prototype. |
| 2371 | LibFunc F; |
| 2372 | if (!TLI->getLibFunc(FDecl: *Fn, F) || F != Func) |
| 2373 | return nullptr; |
| 2374 | |
| 2375 | return Fn; |
| 2376 | } |
| 2377 | |
| 2378 | /// Returns whether the given function is an empty C++ destructor or atexit |
| 2379 | /// handler and can therefore be eliminated. Note that we assume that other |
| 2380 | /// optimization passes have already simplified the code so we simply check for |
| 2381 | /// 'ret'. |
| 2382 | static bool IsEmptyAtExitFunction(const Function &Fn) { |
| 2383 | // FIXME: We could eliminate C++ destructors if they're readonly/readnone and |
| 2384 | // nounwind, but that doesn't seem worth doing. |
| 2385 | if (Fn.isDeclaration()) |
| 2386 | return false; |
| 2387 | |
| 2388 | for (const auto &I : Fn.getEntryBlock()) { |
| 2389 | if (I.isDebugOrPseudoInst()) |
| 2390 | continue; |
| 2391 | if (isa<ReturnInst>(Val: I)) |
| 2392 | return true; |
| 2393 | break; |
| 2394 | } |
| 2395 | return false; |
| 2396 | } |
| 2397 | |
| 2398 | static bool OptimizeEmptyGlobalAtExitDtors(Function *CXAAtExitFn, bool isCXX) { |
| 2399 | /// Itanium C++ ABI p3.3.5: |
| 2400 | /// |
| 2401 | /// After constructing a global (or local static) object, that will require |
| 2402 | /// destruction on exit, a termination function is registered as follows: |
| 2403 | /// |
| 2404 | /// extern "C" int __cxa_atexit ( void (*f)(void *), void *p, void *d ); |
| 2405 | /// |
| 2406 | /// This registration, e.g. __cxa_atexit(f,p,d), is intended to cause the |
| 2407 | /// call f(p) when DSO d is unloaded, before all such termination calls |
| 2408 | /// registered before this one. It returns zero if registration is |
| 2409 | /// successful, nonzero on failure. |
| 2410 | |
| 2411 | // This pass will look for calls to __cxa_atexit or atexit where the function |
| 2412 | // is trivial and remove them. |
| 2413 | bool Changed = false; |
| 2414 | |
| 2415 | for (User *U : llvm::make_early_inc_range(Range: CXAAtExitFn->users())) { |
| 2416 | // We're only interested in calls. Theoretically, we could handle invoke |
| 2417 | // instructions as well, but neither llvm-gcc nor clang generate invokes |
| 2418 | // to __cxa_atexit. |
| 2419 | CallInst *CI = dyn_cast<CallInst>(Val: U); |
| 2420 | if (!CI) |
| 2421 | continue; |
| 2422 | |
| 2423 | Function *DtorFn = |
| 2424 | dyn_cast<Function>(Val: CI->getArgOperand(i: 0)->stripPointerCasts()); |
| 2425 | if (!DtorFn || !IsEmptyAtExitFunction(Fn: *DtorFn)) |
| 2426 | continue; |
| 2427 | |
| 2428 | // Just remove the call. |
| 2429 | CI->replaceAllUsesWith(V: Constant::getNullValue(Ty: CI->getType())); |
| 2430 | CI->eraseFromParent(); |
| 2431 | |
| 2432 | if (isCXX) |
| 2433 | ++NumCXXDtorsRemoved; |
| 2434 | else |
| 2435 | ++NumAtExitRemoved; |
| 2436 | |
| 2437 | Changed |= true; |
| 2438 | } |
| 2439 | |
| 2440 | return Changed; |
| 2441 | } |
| 2442 | |
| 2443 | static Function *hasSideeffectFreeStaticResolution(GlobalIFunc &IF) { |
| 2444 | if (IF.isInterposable()) |
| 2445 | return nullptr; |
| 2446 | |
| 2447 | Function *Resolver = IF.getResolverFunction(); |
| 2448 | if (!Resolver) |
| 2449 | return nullptr; |
| 2450 | |
| 2451 | if (Resolver->isInterposable()) |
| 2452 | return nullptr; |
| 2453 | |
| 2454 | // Only handle functions that have been optimized into a single basic block. |
| 2455 | auto It = Resolver->begin(); |
| 2456 | if (++It != Resolver->end()) |
| 2457 | return nullptr; |
| 2458 | |
| 2459 | BasicBlock &BB = Resolver->getEntryBlock(); |
| 2460 | |
| 2461 | if (any_of(Range&: BB, P: [](Instruction &I) { return I.mayHaveSideEffects(); })) |
| 2462 | return nullptr; |
| 2463 | |
| 2464 | auto *Ret = dyn_cast<ReturnInst>(Val: BB.getTerminator()); |
| 2465 | if (!Ret) |
| 2466 | return nullptr; |
| 2467 | |
| 2468 | return dyn_cast<Function>(Val: Ret->getReturnValue()); |
| 2469 | } |
| 2470 | |
| 2471 | /// Find IFuncs that have resolvers that always point at the same statically |
| 2472 | /// known callee, and replace their callers with a direct call. |
| 2473 | static bool OptimizeStaticIFuncs(Module &M) { |
| 2474 | bool Changed = false; |
| 2475 | for (GlobalIFunc &IF : M.ifuncs()) |
| 2476 | if (Function *Callee = hasSideeffectFreeStaticResolution(IF)) |
| 2477 | if (!IF.use_empty() && |
| 2478 | (!Callee->isDeclaration() || |
| 2479 | none_of(Range: IF.users(), P: [](User *U) { return isa<GlobalAlias>(Val: U); }))) { |
| 2480 | IF.replaceAllUsesWith(V: Callee); |
| 2481 | NumIFuncsResolved++; |
| 2482 | Changed = true; |
| 2483 | } |
| 2484 | return Changed; |
| 2485 | } |
| 2486 | |
| 2487 | static bool |
| 2488 | DeleteDeadIFuncs(Module &M, |
| 2489 | SmallPtrSetImpl<const Comdat *> &NotDiscardableComdats) { |
| 2490 | bool Changed = false; |
| 2491 | for (GlobalIFunc &IF : make_early_inc_range(Range: M.ifuncs())) |
| 2492 | if (deleteIfDead(GV&: IF, NotDiscardableComdats)) { |
| 2493 | NumIFuncsDeleted++; |
| 2494 | Changed = true; |
| 2495 | } |
| 2496 | return Changed; |
| 2497 | } |
| 2498 | |
| 2499 | // Follows the use-def chain of \p V backwards until it finds a Function, |
| 2500 | // in which case it collects in \p Versions. Return true on successful |
| 2501 | // use-def chain traversal, false otherwise. |
| 2502 | static bool |
| 2503 | collectVersions(Value *V, SmallVectorImpl<Function *> &Versions, |
| 2504 | function_ref<TargetTransformInfo &(Function &)> GetTTI) { |
| 2505 | if (auto *F = dyn_cast<Function>(Val: V)) { |
| 2506 | if (!GetTTI(*F).isMultiversionedFunction(F: *F)) |
| 2507 | return false; |
| 2508 | Versions.push_back(Elt: F); |
| 2509 | } else if (auto *Sel = dyn_cast<SelectInst>(Val: V)) { |
| 2510 | if (!collectVersions(V: Sel->getTrueValue(), Versions, GetTTI)) |
| 2511 | return false; |
| 2512 | if (!collectVersions(V: Sel->getFalseValue(), Versions, GetTTI)) |
| 2513 | return false; |
| 2514 | } else if (auto *Phi = dyn_cast<PHINode>(Val: V)) { |
| 2515 | for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) |
| 2516 | if (!collectVersions(V: Phi->getIncomingValue(i: I), Versions, GetTTI)) |
| 2517 | return false; |
| 2518 | } else { |
| 2519 | // Unknown instruction type. Bail. |
| 2520 | return false; |
| 2521 | } |
| 2522 | return true; |
| 2523 | } |
| 2524 | |
| 2525 | // Try to statically resolve calls to versioned functions when possible. First |
| 2526 | // we identify the function versions which are associated with an IFUNC symbol. |
| 2527 | // We do that by examining the resolver function of the IFUNC. Once we have |
| 2528 | // collected all the function versions, we sort them in decreasing priority |
| 2529 | // order. This is necessary for determining the most suitable callee version |
| 2530 | // for each caller version. We then collect all the callsites to versioned |
| 2531 | // functions. The static resolution is performed by comparing the feature sets |
| 2532 | // between callers and callees. Specifically: |
| 2533 | // * Start a walk over caller and callee lists simultaneously in order of |
| 2534 | // decreasing priority. |
| 2535 | // * Statically resolve calls from the current caller to the current callee, |
| 2536 | // iff the caller feature bits are a superset of the callee feature bits. |
| 2537 | // * For FMV callers, as long as the caller feature bits are a subset of the |
| 2538 | // callee feature bits, advance to the next callee. This effectively prevents |
| 2539 | // considering the current callee as a candidate for static resolution by |
| 2540 | // following callers (explanation: preceding callers would not have been |
| 2541 | // selected in a hypothetical runtime execution). |
| 2542 | // * Advance to the next caller. |
| 2543 | // |
| 2544 | // Presentation in EuroLLVM2025: |
| 2545 | // https://www.youtube.com/watch?v=k54MFimPz-A&t=867s |
| 2546 | static bool OptimizeNonTrivialIFuncs( |
| 2547 | Module &M, function_ref<TargetTransformInfo &(Function &)> GetTTI) { |
| 2548 | bool Changed = false; |
| 2549 | |
| 2550 | // Map containing the feature bits for a given function. |
| 2551 | DenseMap<Function *, APInt> FeatureMask; |
| 2552 | // Map containing the priority bits for a given function. |
| 2553 | DenseMap<Function *, APInt> PriorityMask; |
| 2554 | // Map containing all the function versions corresponding to an IFunc symbol. |
| 2555 | DenseMap<GlobalIFunc *, SmallVector<Function *>> VersionedFuncs; |
| 2556 | // Map containing the IFunc symbol a function is version of. |
| 2557 | DenseMap<Function *, GlobalIFunc *> VersionOf; |
| 2558 | // List of all the interesting IFuncs found in the module. |
| 2559 | SmallVector<GlobalIFunc *> IFuncs; |
| 2560 | |
| 2561 | for (GlobalIFunc &IF : M.ifuncs()) { |
| 2562 | LLVM_DEBUG(dbgs() << "Examining IFUNC " << IF.getName() << "\n" ); |
| 2563 | |
| 2564 | if (IF.isInterposable()) |
| 2565 | continue; |
| 2566 | |
| 2567 | Function *Resolver = IF.getResolverFunction(); |
| 2568 | if (!Resolver) |
| 2569 | continue; |
| 2570 | |
| 2571 | if (Resolver->isInterposable()) |
| 2572 | continue; |
| 2573 | |
| 2574 | SmallVector<Function *> Versions; |
| 2575 | // Discover the versioned functions. |
| 2576 | if (any_of(Range&: *Resolver, P: [&](BasicBlock &BB) { |
| 2577 | if (auto *Ret = dyn_cast_or_null<ReturnInst>(Val: BB.getTerminator())) |
| 2578 | if (!collectVersions(V: Ret->getReturnValue(), Versions, GetTTI)) |
| 2579 | return true; |
| 2580 | return false; |
| 2581 | })) |
| 2582 | continue; |
| 2583 | |
| 2584 | if (Versions.empty()) |
| 2585 | continue; |
| 2586 | |
| 2587 | for (Function *V : Versions) { |
| 2588 | VersionOf.insert(KV: {V, &IF}); |
| 2589 | auto [FeatIt, FeatInserted] = FeatureMask.try_emplace(Key: V); |
| 2590 | if (FeatInserted) |
| 2591 | FeatIt->second = GetTTI(*V).getFeatureMask(F: *V); |
| 2592 | auto [PriorIt, PriorInserted] = PriorityMask.try_emplace(Key: V); |
| 2593 | if (PriorInserted) |
| 2594 | PriorIt->second = GetTTI(*V).getPriorityMask(F: *V); |
| 2595 | } |
| 2596 | |
| 2597 | // Sort function versions in decreasing priority order. |
| 2598 | sort(C&: Versions, Comp: [&](auto *LHS, auto *RHS) { |
| 2599 | return PriorityMask[LHS].ugt(PriorityMask[RHS]); |
| 2600 | }); |
| 2601 | |
| 2602 | IFuncs.push_back(Elt: &IF); |
| 2603 | VersionedFuncs.try_emplace(Key: &IF, Args: std::move(Versions)); |
| 2604 | } |
| 2605 | |
| 2606 | for (GlobalIFunc *CalleeIF : IFuncs) { |
| 2607 | SmallVector<Function *> NonFMVCallers; |
| 2608 | DenseSet<GlobalIFunc *> CallerIFuncs; |
| 2609 | DenseMap<Function *, SmallVector<CallBase *>> CallSites; |
| 2610 | |
| 2611 | // Find the callsites. |
| 2612 | for (User *U : CalleeIF->users()) { |
| 2613 | if (auto *CB = dyn_cast<CallBase>(Val: U)) { |
| 2614 | if (CB->getCalledOperand() == CalleeIF) { |
| 2615 | Function *Caller = CB->getFunction(); |
| 2616 | GlobalIFunc *CallerIF = nullptr; |
| 2617 | TargetTransformInfo &TTI = GetTTI(*Caller); |
| 2618 | bool CallerIsFMV = TTI.isMultiversionedFunction(F: *Caller); |
| 2619 | // The caller is a version of a known IFunc. |
| 2620 | if (auto It = VersionOf.find(Val: Caller); It != VersionOf.end()) |
| 2621 | CallerIF = It->second; |
| 2622 | else if (!CallerIsFMV && OptimizeNonFMVCallers) { |
| 2623 | // The caller is non-FMV. |
| 2624 | auto [It, Inserted] = FeatureMask.try_emplace(Key: Caller); |
| 2625 | if (Inserted) |
| 2626 | It->second = TTI.getFeatureMask(F: *Caller); |
| 2627 | } else |
| 2628 | // The caller is none of the above, skip. |
| 2629 | continue; |
| 2630 | auto [It, Inserted] = CallSites.try_emplace(Key: Caller); |
| 2631 | if (Inserted) { |
| 2632 | if (CallerIsFMV) |
| 2633 | CallerIFuncs.insert(V: CallerIF); |
| 2634 | else |
| 2635 | NonFMVCallers.push_back(Elt: Caller); |
| 2636 | } |
| 2637 | It->second.push_back(Elt: CB); |
| 2638 | } |
| 2639 | } |
| 2640 | } |
| 2641 | |
| 2642 | if (CallSites.empty()) |
| 2643 | continue; |
| 2644 | |
| 2645 | LLVM_DEBUG(dbgs() << "Statically resolving calls to function " |
| 2646 | << CalleeIF->getResolverFunction()->getName() << "\n" ); |
| 2647 | |
| 2648 | // The complexity of this algorithm is linear: O(NumCallers + NumCallees) |
| 2649 | // if NumCallers > MaxIFuncVersions || NumCallees > MaxIFuncVersions, |
| 2650 | // otherwise it is cubic: O((NumCallers ^ 2) x NumCallees). |
| 2651 | auto staticallyResolveCalls = [&](ArrayRef<Function *> Callers, |
| 2652 | ArrayRef<Function *> Callees, |
| 2653 | bool CallerIsFMV) { |
| 2654 | bool AllowExpensiveChecks = CallerIsFMV && |
| 2655 | Callers.size() <= MaxIFuncVersions && |
| 2656 | Callees.size() <= MaxIFuncVersions; |
| 2657 | // Index to the highest callee candidate. |
| 2658 | unsigned J = 0; |
| 2659 | |
| 2660 | for (unsigned I = 0, E = Callers.size(); I < E; ++I) { |
| 2661 | // There are no callee candidates left. |
| 2662 | if (J == Callees.size()) |
| 2663 | break; |
| 2664 | |
| 2665 | Function *Caller = Callers[I]; |
| 2666 | APInt CallerBits = FeatureMask[Caller]; |
| 2667 | |
| 2668 | // Compare the feature bits of the best callee candidate with all the |
| 2669 | // caller versions preceeding the current one. For each prior caller |
| 2670 | // discard feature bits that are known to be available in the current |
| 2671 | // caller. As long as the known missing feature bits are a subset of the |
| 2672 | // callee feature bits, advance to the next callee and start over. |
| 2673 | auto eliminateAvailableFeatures = [&](unsigned BestCandidate) { |
| 2674 | unsigned K = 0; |
| 2675 | while (K < I && BestCandidate < Callees.size()) { |
| 2676 | APInt MissingBits = FeatureMask[Callers[K]] & ~CallerBits; |
| 2677 | if (MissingBits.isSubsetOf(RHS: FeatureMask[Callees[BestCandidate]])) { |
| 2678 | ++BestCandidate; |
| 2679 | // Start over. |
| 2680 | K = 0; |
| 2681 | } else |
| 2682 | ++K; |
| 2683 | } |
| 2684 | return BestCandidate; |
| 2685 | }; |
| 2686 | |
| 2687 | unsigned BestCandidate = |
| 2688 | AllowExpensiveChecks ? eliminateAvailableFeatures(J) : J; |
| 2689 | // No callee candidate was found for this caller. |
| 2690 | if (BestCandidate == Callees.size()) |
| 2691 | continue; |
| 2692 | |
| 2693 | LLVM_DEBUG(dbgs() << " Examining " |
| 2694 | << (CallerIsFMV ? "FMV" : "regular" ) << " caller " |
| 2695 | << Caller->getName() << "\n" ); |
| 2696 | |
| 2697 | Function *Callee = Callees[BestCandidate]; |
| 2698 | APInt CalleeBits = FeatureMask[Callee]; |
| 2699 | |
| 2700 | // Statically resolve calls from the current caller to the current |
| 2701 | // callee, iff the caller feature bits are a superset of the callee |
| 2702 | // feature bits. |
| 2703 | if (CalleeBits.isSubsetOf(RHS: CallerBits)) { |
| 2704 | // Not all caller versions are necessarily users of the callee IFUNC. |
| 2705 | if (auto It = CallSites.find(Val: Caller); It != CallSites.end()) { |
| 2706 | for (CallBase *CS : It->second) { |
| 2707 | LLVM_DEBUG(dbgs() << " Redirecting call " << Caller->getName() |
| 2708 | << " -> " << Callee->getName() << "\n" ); |
| 2709 | CS->setCalledOperand(Callee); |
| 2710 | } |
| 2711 | Changed = true; |
| 2712 | } |
| 2713 | } |
| 2714 | |
| 2715 | // Nothing else to do about non-FMV callers. |
| 2716 | if (!CallerIsFMV) |
| 2717 | continue; |
| 2718 | |
| 2719 | // For FMV callers, as long as the caller feature bits are a subset of |
| 2720 | // the callee feature bits, advance to the next callee. This effectively |
| 2721 | // prevents considering the current callee as a candidate for static |
| 2722 | // resolution by following callers. |
| 2723 | while (CallerBits.isSubsetOf(RHS: FeatureMask[Callees[J]]) && |
| 2724 | ++J < Callees.size()) |
| 2725 | ; |
| 2726 | } |
| 2727 | }; |
| 2728 | |
| 2729 | auto &Callees = VersionedFuncs[CalleeIF]; |
| 2730 | |
| 2731 | // Optimize non-FMV calls. |
| 2732 | if (OptimizeNonFMVCallers) |
| 2733 | staticallyResolveCalls(NonFMVCallers, Callees, /*CallerIsFMV=*/false); |
| 2734 | |
| 2735 | // Optimize FMV calls. |
| 2736 | for (GlobalIFunc *CallerIF : CallerIFuncs) { |
| 2737 | auto &Callers = VersionedFuncs[CallerIF]; |
| 2738 | staticallyResolveCalls(Callers, Callees, /*CallerIsFMV=*/true); |
| 2739 | } |
| 2740 | |
| 2741 | if (CalleeIF->use_empty() || |
| 2742 | all_of(Range: CalleeIF->users(), P: [](User *U) { return isa<GlobalAlias>(Val: U); })) |
| 2743 | NumIFuncsResolved++; |
| 2744 | } |
| 2745 | return Changed; |
| 2746 | } |
| 2747 | |
| 2748 | static bool |
| 2749 | optimizeGlobalsInModule(Module &M, const DataLayout &DL, |
| 2750 | function_ref<TargetLibraryInfo &(Function &)> GetTLI, |
| 2751 | function_ref<TargetTransformInfo &(Function &)> GetTTI, |
| 2752 | function_ref<BlockFrequencyInfo &(Function &)> GetBFI, |
| 2753 | function_ref<DominatorTree &(Function &)> LookupDomTree, |
| 2754 | function_ref<void(Function &F)> ChangedCFGCallback, |
| 2755 | function_ref<void(Function &F)> DeleteFnCallback) { |
| 2756 | SmallPtrSet<const Comdat *, 8> NotDiscardableComdats; |
| 2757 | bool Changed = false; |
| 2758 | bool LocalChange = true; |
| 2759 | std::optional<uint32_t> FirstNotFullyEvaluatedPriority; |
| 2760 | |
| 2761 | while (LocalChange) { |
| 2762 | LocalChange = false; |
| 2763 | |
| 2764 | NotDiscardableComdats.clear(); |
| 2765 | for (const GlobalVariable &GV : M.globals()) |
| 2766 | if (const Comdat *C = GV.getComdat()) |
| 2767 | if (!GV.isDiscardableIfUnused() || !GV.use_empty()) |
| 2768 | NotDiscardableComdats.insert(Ptr: C); |
| 2769 | for (Function &F : M) |
| 2770 | if (const Comdat *C = F.getComdat()) |
| 2771 | if (!F.isDefTriviallyDead()) |
| 2772 | NotDiscardableComdats.insert(Ptr: C); |
| 2773 | for (GlobalAlias &GA : M.aliases()) |
| 2774 | if (const Comdat *C = GA.getComdat()) |
| 2775 | if (!GA.isDiscardableIfUnused() || !GA.use_empty()) |
| 2776 | NotDiscardableComdats.insert(Ptr: C); |
| 2777 | |
| 2778 | // Delete functions that are trivially dead, ccc -> fastcc |
| 2779 | LocalChange |= OptimizeFunctions(M, GetTLI, GetTTI, GetBFI, LookupDomTree, |
| 2780 | NotDiscardableComdats, ChangedCFGCallback, |
| 2781 | DeleteFnCallback); |
| 2782 | |
| 2783 | // Optimize global_ctors list. |
| 2784 | LocalChange |= |
| 2785 | optimizeGlobalCtorsList(M, ShouldRemove: [&](uint32_t Priority, Function *F) { |
| 2786 | if (FirstNotFullyEvaluatedPriority && |
| 2787 | *FirstNotFullyEvaluatedPriority != Priority) |
| 2788 | return false; |
| 2789 | bool Evaluated = EvaluateStaticConstructor(F, DL, TLI: &GetTLI(*F)); |
| 2790 | if (!Evaluated) |
| 2791 | FirstNotFullyEvaluatedPriority = Priority; |
| 2792 | return Evaluated; |
| 2793 | }); |
| 2794 | |
| 2795 | // Optimize non-address-taken globals. |
| 2796 | LocalChange |= OptimizeGlobalVars(M, GetTTI, GetTLI, LookupDomTree, |
| 2797 | NotDiscardableComdats); |
| 2798 | |
| 2799 | // Resolve aliases, when possible. |
| 2800 | LocalChange |= OptimizeGlobalAliases(M, NotDiscardableComdats); |
| 2801 | |
| 2802 | // Try to remove trivial global destructors if they are not removed |
| 2803 | // already. |
| 2804 | if (Function *CXAAtExitFn = |
| 2805 | FindAtExitLibFunc(M, GetTLI, Func: LibFunc_cxa_atexit)) |
| 2806 | LocalChange |= OptimizeEmptyGlobalAtExitDtors(CXAAtExitFn, isCXX: true); |
| 2807 | |
| 2808 | if (Function *AtExitFn = FindAtExitLibFunc(M, GetTLI, Func: LibFunc_atexit)) |
| 2809 | LocalChange |= OptimizeEmptyGlobalAtExitDtors(CXAAtExitFn: AtExitFn, isCXX: false); |
| 2810 | |
| 2811 | // Optimize IFuncs whose callee's are statically known. |
| 2812 | LocalChange |= OptimizeStaticIFuncs(M); |
| 2813 | |
| 2814 | // Optimize IFuncs based on the target features of the caller. |
| 2815 | LocalChange |= OptimizeNonTrivialIFuncs(M, GetTTI); |
| 2816 | |
| 2817 | // Remove any IFuncs that are now dead. |
| 2818 | LocalChange |= DeleteDeadIFuncs(M, NotDiscardableComdats); |
| 2819 | |
| 2820 | Changed |= LocalChange; |
| 2821 | } |
| 2822 | |
| 2823 | // TODO: Move all global ctors functions to the end of the module for code |
| 2824 | // layout. |
| 2825 | |
| 2826 | return Changed; |
| 2827 | } |
| 2828 | |
| 2829 | PreservedAnalyses GlobalOptPass::run(Module &M, ModuleAnalysisManager &AM) { |
| 2830 | auto &DL = M.getDataLayout(); |
| 2831 | auto &FAM = |
| 2832 | AM.getResult<FunctionAnalysisManagerModuleProxy>(IR&: M).getManager(); |
| 2833 | auto LookupDomTree = [&FAM](Function &F) -> DominatorTree &{ |
| 2834 | return FAM.getResult<DominatorTreeAnalysis>(IR&: F); |
| 2835 | }; |
| 2836 | auto GetTLI = [&FAM](Function &F) -> TargetLibraryInfo & { |
| 2837 | return FAM.getResult<TargetLibraryAnalysis>(IR&: F); |
| 2838 | }; |
| 2839 | auto GetTTI = [&FAM](Function &F) -> TargetTransformInfo & { |
| 2840 | return FAM.getResult<TargetIRAnalysis>(IR&: F); |
| 2841 | }; |
| 2842 | |
| 2843 | auto GetBFI = [&FAM](Function &F) -> BlockFrequencyInfo & { |
| 2844 | return FAM.getResult<BlockFrequencyAnalysis>(IR&: F); |
| 2845 | }; |
| 2846 | auto ChangedCFGCallback = [&FAM](Function &F) { |
| 2847 | FAM.invalidate(IR&: F, PA: PreservedAnalyses::none()); |
| 2848 | }; |
| 2849 | auto DeleteFnCallback = [&FAM](Function &F) { FAM.clear(IR&: F, Name: F.getName()); }; |
| 2850 | |
| 2851 | if (!optimizeGlobalsInModule(M, DL, GetTLI, GetTTI, GetBFI, LookupDomTree, |
| 2852 | ChangedCFGCallback, DeleteFnCallback)) |
| 2853 | return PreservedAnalyses::all(); |
| 2854 | |
| 2855 | PreservedAnalyses PA = PreservedAnalyses::none(); |
| 2856 | // We made sure to clear analyses for deleted functions. |
| 2857 | PA.preserve<FunctionAnalysisManagerModuleProxy>(); |
| 2858 | // The only place we modify the CFG is when calling |
| 2859 | // removeUnreachableBlocks(), but there we make sure to invalidate analyses |
| 2860 | // for modified functions. |
| 2861 | PA.preserveSet<CFGAnalyses>(); |
| 2862 | return PA; |
| 2863 | } |
| 2864 | |